CN108288161A

CN108288161A - The method and system of prediction result are provided based on machine learning

Info

Publication number: CN108288161A
Application number: CN201710016236.8A
Authority: CN
Inventors: 黄晶; 陈雨强
Original assignee: 4Paradigm Beijing Technology Co Ltd
Current assignee: 4Paradigm Beijing Technology Co Ltd
Priority date: 2017-01-10
Filing date: 2017-01-10
Publication date: 2018-07-17

Abstract

There is provided it is a kind of providing the method and system of prediction result based on machine learning, the method includes：(A) attribute information of sample to be predicted and the attribute information of the history sample occurred before sample to be predicted are obtained；(B) result information of the history sample about forecasting problem is obtained, wherein for the history sample for not having the legitimate reading about forecasting problem among the history sample, using the prediction result of history sample as the result information of history sample；(C) result information of the attribute information of the sample to be predicted based on acquisition, the attribute information of the history sample and the history sample generates the forecast sample of sample to be predicted；(D) using the prediction model trained based on machine learning techniques, the prediction result of sample to be predicted is provided for the forecast sample of sample to be predicted.According to described method and system, it can suitably be directed to sample to be predicted with reference to both history sample and sample to be predicted and provide prediction result.

Description

The method and system of prediction result are provided based on machine learning

Technical field

All things considered of the present invention is related to artificial intelligence field, is provided based on machine learning more specifically to one kind Method and system of the sample to be predicted about the prediction result of forecasting problem.

Background technology

In practice, in order to provide prediction result of the sample to be predicted about forecasting problem based on machine learning techniques, Other than needing the attribute information of sample itself to be predicted, usually also need to obtain relevant historical information, that is, to be predicted The attribute information of the history sample occurred before sample.For example, if it is desired to predicting that user's is current using machine learning model Whether transaction is fraudulent trading, it is also necessary to the case where obtaining the historical trading of user, and according to current transaction and historical trading It compares to provide prediction result.

However, the attribute information of history sample is in use, it is easy to appear various problems in many cases,.For example, In the example of above-mentioned fraudulent trading, in reality it occur frequently that the phenomenon that continuous fraudulent trading (by taking credit card steals brush as an example, if The first stroke is stolen brush and is not found, then a lot of robber's brushes can occur in succession), at this point, being mixed into abnormal sample in history sample, cause Relativity between history sample and current sample can not be effectively reflected the comparison between normal sample and abnormal sample Relationship so that model can not effectively work.For example, it is assumed that machine learning model is for predicting credit card fraud transaction, model Positive sample to correspond to current sample be abnormal sample, and it is normal sample that the negative sample of model, which corresponds to current sample,.Into one Step ground, it is assumed that the place of swiping the card in the swipe the card place and current transaction of a upper transaction involved in the sample characteristics of model, then For as positive sample twice in succession steal brush among second steal brush for, model be difficult by the sample learning to how Accurate Prediction steals brush.Particularly, if certain credit card user people is in China, and his credit card has brushed two in the U.S. by continuous robber It is secondary, then stealing brush for second and stealing brush for the first time equally all as the positive sample of model.However, corresponding for stealing brush with second Positive sample for, upper brush card place is the U.S., and the place of current brush card is also in the U.S., in this sample Under, model can be easy to tend to think above-mentioned second in sample that the U.S. swipes the card (that is, the in the U.S. swipes the card twice in succession It is secondary to swipe the card) it is easy to happen fraud, and this is not inconsistent with common sense, it can be seen that, the model trained in this manner is difficult Effectively to predict fraudulent trading.

In view of the above-mentioned problems, can consider that qualified history sample is used only, for example, above-mentioned about fraudulent trading In example, historical fraudulent trading is weeded out using only historical arm's length dealing.However, executing prediction at that time (for example, in scene of online Prediction) leads to not sieve from Recent Activity due to the stateful transaction still non-availability of Recent Activity Select arm's length dealing；On the other hand, Recent Activity be the key that again can help to judge currently to merchandise whether be fraudulent trading because Therefore element, the effect that model can be seriously affected if directly neglecting the unknown all Recent Activities of stateful transaction are used only The scheme of qualified history sample feasibility in reality is poor, it is difficult to obtain effective prediction result.

In conclusion when in face of the particular problem of machine learning, need to be improved from modelling angle, to have The computing resource (for example, the limitation of hardware resource in terms of the capacity of processing data and speed) and/or data resource (example of limit Such as, lack enough training samples for training machine learning model) under effectively solve the problems, such as that history sample state is unknown, And then ensure the prediction effect of machine learning model.

Invention content

Exemplary embodiment of the present invention is intended to overcome the existing prediction scheme based on machine learning model to be difficult to effectively The defect learnt from history sample.

Exemplary embodiment according to the present invention provides one kind based on machine learning to provide sample to be predicted about prediction The method of the prediction result of problem, including：(A) it obtains the attribute information of sample to be predicted and occurs before sample to be predicted History sample attribute information；(B) result information of the history sample about forecasting problem is obtained, wherein for described The history sample for not having the legitimate reading about forecasting problem among history sample, using the prediction result of history sample as going through The result information of history sample；(C) attribute information of the sample to be predicted based on acquisition, the history sample attribute information and The result information of the history sample generates the forecast sample of sample to be predicted；And (D) is utilized and is based on machine learning techniques The prediction model trained provides the prediction result of sample to be predicted for the forecast sample of sample to be predicted.

Optionally, in the method, in step (B), the confidence level of the result information of history sample is also obtained, and And in step (C), the attribute information of the sample to be predicted based on acquisition, the attribute information of the history sample, the history The confidence level of the result information of the result information of sample and the history sample generates the forecast sample of sample to be predicted.

Optionally, in the method, in step (B), the prediction result of the history sample is by the prediction model Or previous model corresponding with the last round of iteration of the prediction model provides.

Optionally, in the method, in step (B), for the history sample with legitimate reading, by history sample Result information of the legitimate reading as history sample.

Optionally, in the method, there is the prediction model following training process to be pressed in the training process According to the mode consistent with forecast sample, the legitimate reading or prediction result for the history sample that training sample is based on are as described in The result information of history sample, wherein the prediction result of the history sample that the training sample is based on is by currently training Prediction model provides.

Optionally, in the method, in the training process, the prediction model is iterated for training sample Training so that the prediction result for the history sample that training sample is based on is constantly updated with iteration.

Optionally, the method provides prediction result of the sample to be predicted about forecasting problem online.

Optionally, in the method, the pre- test sample of sample to be predicted is generated at least one of in the following manner This feature：(C1) it is sieved according to the confidence level of the result information of the history sample and the result information of the history sample It selects at least part history sample, and the attribute information based at least part history sample filtered out and waits for pre- The attribute information of test sample example generates the feature of the forecast sample of sample to be predicted；(C2) believed according to the result of the history sample The confidence level of the result information of breath and the history sample is weighted the respective attributes information of the history sample, and base The attribute information of history sample after weighting and the attribute information of sample to be predicted generate the pre- test sample of sample to be predicted This feature；And (C3) is based respectively on the attribute information of sample to be predicted, the attribute information of the history sample, the history The confidence level of the result information of the result information of sample and the history sample generates the forecast sample of sample to be predicted Feature.

Optionally, in the method, sample to be predicted corresponds to current transaction, and history sample corresponds to merchandises currently The previous transaction of the predetermined quantity occurred before and/or the previous transaction occurred in the predetermined amount of time before current transaction, Forecasting problem is whether relationship trading is fraudulent trading.

Optionally, in the method, in step (B), for the history sample without legitimate reading, based on prediction The algorithm of model obtains the confidence level of the prediction result of history sample independently of the algorithm of prediction model, as history sample Result information confidence level；For the history sample with legitimate reading, the confidence level of the legitimate reading of history sample is set It is set to the preset value for indicating high confidence level, the confidence level of the result information as history sample.

Optionally, the method further includes：(E) legitimate reading of the sample to be predicted about forecasting problem is received, wherein institute Legitimate reading is stated to be used to train the prediction model together with corresponding sample to be predicted.

Another exemplary according to the present invention is implemented, and provides one kind based on machine learning to provide sample to be predicted about pre- The system of the prediction result of survey problem, including：Attribute information acquisition device, attribute information for obtaining sample to be predicted and The attribute information of the history sample occurred before sample to be predicted；Result information acquisition device, for obtaining the history sample Result information of the example about forecasting problem, wherein for the true knot not having among the history sample about forecasting problem The history sample of fruit, result information acquisition device is using the prediction result of history sample as the result information of history sample；Sample Generating means for the attribute information of the sample to be predicted based on acquisition, the attribute information of the history sample and described are gone through The result information of history sample generates the forecast sample of sample to be predicted；And prediction result provides device, is based on for utilizing The prediction model that machine learning techniques train provides the prediction knot of sample to be predicted for the forecast sample of sample to be predicted Fruit.

Optionally, in the system, result information acquisition device also obtains the confidence level of the result information of history sample, Also, the attribute information of to be predicted sample of the sample generating means based on acquisition, described is gone through the attribute information of the history sample The confidence level of the result information of the result information of history sample and the history sample generates the forecast sample of sample to be predicted.

Optionally, in the system, the prediction result of the history sample obtained by result information acquisition device by The prediction model or previous model corresponding with the last round of iteration of the prediction model provide.

Optionally, in the system, for the history sample with legitimate reading, result information acquisition device is by history Result information of the legitimate reading of sample as history sample.

Optionally, in the system, there is the prediction model following training process to be pressed in the training process According to the mode consistent with forecast sample, the legitimate reading or prediction result for the history sample that training sample is based on are as described in The result information of history sample, wherein the prediction result of the history sample that the training sample is based on is by currently training Prediction model provides.

Optionally, in the system, in the training process, the prediction model is iterated for training sample Training so that the prediction result for the history sample that training sample is based on is constantly updated with iteration.

Optionally, in the system, the system provides prediction result of the sample to be predicted about forecasting problem online.

Optionally, in the system, sample generating means are at least one of in the following manner to be predicted to generate The feature of the forecast sample of sample：According to setting for the result information of the result information of the history sample and the history sample Reliability filters out at least part history sample, and based on the attribute information of at least part history sample filtered out And the attribute information of sample to be predicted generates the feature of the forecast sample of sample to be predicted；According to the knot of the history sample The confidence level of the result information of fruit information and the history sample is weighted the respective attributes information of the history sample, And the pre- of sample to be predicted is generated based on the attribute information of the attribute information of the history sample after weighting and sample to be predicted The feature of test sample sheet；And it is based respectively on the attribute information of sample to be predicted, the attribute information of the history sample, the history The confidence level of the result information of the result information of sample and the history sample generates the forecast sample of sample to be predicted Feature.

Optionally, in the system, sample to be predicted corresponds to current transaction, and history sample corresponds to merchandises currently The previous transaction of the predetermined quantity occurred before and/or the previous transaction occurred in the predetermined amount of time before current transaction, Forecasting problem is whether relationship trading is fraudulent trading.

Optionally, in the system, for the history sample without legitimate reading, result information acquisition device is based on The algorithm of prediction model obtains the confidence level of the prediction result of history sample independently of the algorithm of prediction model, as history The confidence level of the result information of sample；For the history sample with legitimate reading, result information acquisition device is by history sample The confidence level of legitimate reading be set as indicating the preset value of high confidence level, the confidence of the result information as history sample Degree.

Optionally, the system also includes：Feedback device, for receiving true knot of the sample to be predicted about forecasting problem Fruit, wherein the legitimate reading be used to train the prediction model together with corresponding sample to be predicted.

Exemplary embodiment according to the present invention provides one kind based on machine learning to provide sample to be predicted about prediction The computer-readable medium of the prediction result of problem, wherein record is useful for executing following on the computer-readable medium The computer program of step：(A) the history sample for obtaining the attribute information of sample to be predicted and occurring before sample to be predicted The attribute information of example；(B) result information of the history sample about forecasting problem is obtained, wherein be directed to the history sample Among do not have about forecasting problem legitimate reading history sample, using the prediction result of history sample as history sample Result information；(C) attribute information of the sample to be predicted based on acquisition, the attribute information of the history sample and the history The result information of sample generates the forecast sample of sample to be predicted；And (D) based on machine learning techniques using being trained Prediction model provides the prediction result of sample to be predicted for the forecast sample of sample to be predicted.

Optionally, in the computer-readable medium, in step (B), the result information of history sample is also obtained Confidence level, also, in step (C), the attribute information of the sample to be predicted based on acquisition, the history sample attribute letter Breath, the result information of the history sample and the history sample the confidence level of result information generate sample to be predicted Forecast sample.

Optionally, in the computer-readable medium, in step (B), the prediction result of the history sample is by institute It states prediction model or previous model corresponding with the last round of iteration of the prediction model provides.

Optionally, in the computer-readable medium, in step (B), for the history sample with legitimate reading, Using the legitimate reading of history sample as the result information of history sample.

Optionally, in the computer-readable medium, the prediction model has following training process, in the training In the process, according to the mode consistent with forecast sample, the legitimate reading or prediction knot of the history sample that training sample is based on Result information of the fruit as the history sample, wherein the prediction result for the history sample that the training sample is based on is by working as Before the prediction model that trains provide.

Optionally, in the computer-readable medium, in the training process, the prediction model is directed to training sample Originally it is iterated training so that the prediction result for the history sample that training sample is based on is constantly updated with iteration.

Optionally, in the computer-readable medium, the computer program is performed to provide sample to be predicted online Prediction result of the example about forecasting problem.

Optionally, in the computer-readable medium, sample to be predicted is generated at least one of in the following manner The feature of the forecast sample of example：(C1) according to the result information of the result information of the history sample and the history sample Confidence level is believed to filter out at least part history sample based on the attribute of at least part history sample filtered out The attribute information of breath and sample to be predicted generates the feature of the forecast sample of sample to be predicted；(C2) according to the history sample The confidence level of the result information of example and the result information of the history sample to the respective attributes information of the history sample into Row weighting, and sample to be predicted is generated based on the attribute information of the attribute information of the history sample after weighting and sample to be predicted The feature of the forecast sample of example；And (C3) is based respectively on the attribute letter of the attribute information of sample to be predicted, the history sample Breath, the result information of the history sample and the history sample the confidence level of result information generate sample to be predicted The feature of forecast sample.

Optionally, in the computer-readable medium, sample to be predicted corresponds to current transaction, and history sample corresponds to The previous transaction of the predetermined quantity occurred before current transaction and/or occur in the predetermined amount of time before current transaction Previously transaction, forecasting problem were whether relationship trading is fraudulent trading.

Optionally, in the computer-readable medium, in step (B), for the history sample without legitimate reading Example, algorithm based on prediction model or obtains the confidence level of the prediction result of history sample independently of the algorithm of prediction model, The confidence level of result information as history sample；For the history sample with legitimate reading, by the true knot of history sample The confidence level of fruit is set as indicating the preset value of high confidence level, the confidence level of the result information as history sample.

Optionally, in the computer-readable medium, the computer program also executes following steps：(E) it receives and waits for Predict legitimate reading of the sample about forecasting problem, wherein the legitimate reading be used to instruct together with corresponding sample to be predicted Practice the prediction model.

Another exemplary according to the present invention is implemented, and provides one kind based on machine learning to provide sample to be predicted about pre- It is executable to be stored with computer in storage unit for the computing device of the prediction result of survey problem, including storage unit and processor Instruction set executes following step when set of computer-executable instructions conjunction is executed by the processor：(A) it obtains and waits for Predict the attribute information of sample and the attribute information of the history sample occurred before sample to be predicted；(B) it is gone through described in obtaining Result information of the history sample about forecasting problem, wherein do not have about the true of forecasting problem among the history sample The history sample of real result, using the prediction result of history sample as the result information of history sample；(C) based on acquisition wait for it is pre- The result information of the attribute information of test sample example, the attribute information of the history sample and the history sample is to be predicted to generate The forecast sample of sample；And (D) utilizes the prediction model trained based on machine learning techniques, for the pre- of sample to be predicted Test sample provided the prediction result of sample to be predicted originally.

Optionally, in the computing device, the confidence level of the result information of history sample is also obtained, also, in step (C) in, the attribute information of the sample to be predicted based on acquisition, the attribute information of the history sample, the history sample knot The confidence level of the result information of fruit information and the history sample generates the forecast sample of sample to be predicted.

Optionally, in the computing device, in step (B), the prediction result of the history sample is by the prediction Model or previous model corresponding with the last round of iteration of the prediction model provide.

Optionally, in the computing device, in step (B), for the history sample with legitimate reading, by history Result information of the legitimate reading of sample as history sample.

Optionally, in the computing device, the prediction model has following training process, in the training process In, according to the mode consistent with forecast sample, the legitimate reading or prediction result of the history sample that training sample is based on are made For the result information of the history sample, wherein the prediction result for the history sample that the training sample is based on is by currently instructing The prediction model practised provides.

Optionally, in the computing device, in the training process, the prediction model is carried out for training sample Repetitive exercise so that the prediction result for the history sample that training sample is based on is constantly updated with iteration.

Optionally, in the computing device, the computing device provides sample to be predicted about forecasting problem online Prediction result.

Optionally, in the computing device, the pre- of sample to be predicted is generated at least one of in the following manner The feature of test sample sheet：(C1) according to the confidence level of the result information of the history sample and the result information of the history sample Filter out at least part history sample, and the attribute information based at least part history sample filtered out and The attribute information of sample to be predicted generates the feature of the forecast sample of sample to be predicted；(C2) according to the knot of the history sample The confidence level of the result information of fruit information and the history sample is weighted the respective attributes information of the history sample, And the pre- of sample to be predicted is generated based on the attribute information of the attribute information of the history sample after weighting and sample to be predicted The feature of test sample sheet；And (C3) is based respectively on the attribute information, described of the attribute information of sample to be predicted, the history sample The confidence level of the result information of the result information of history sample and the history sample generates the pre- test sample of sample to be predicted This feature.

Optionally, in the computing device, sample to be predicted corresponds to current transaction, and history sample corresponds to current The previous transaction for the predetermined quantity that transaction occurs before and/or the previous friendship occurred in the predetermined amount of time before current transaction Easily, forecasting problem is whether relationship trading is fraudulent trading.

Optionally, it in the computing device, in step (B), for the history sample without legitimate reading, is based on The algorithm of prediction model obtains the confidence level of the prediction result of history sample independently of the algorithm of prediction model, as history The confidence level of the result information of sample；For the history sample with legitimate reading, by the confidence of the legitimate reading of history sample Degree is set as indicating the preset value of high confidence level, the confidence level of the result information as history sample.

Optionally, computing device further includes：(E) legitimate reading of the sample to be predicted about forecasting problem is received, wherein institute Legitimate reading is stated to be used to train the prediction model together with corresponding sample to be predicted.

The method and system of prediction result are provided based on machine learning according to an exemplary embodiment of the present invention, can be had Effect ground obtains result information of the history sample about forecasting problem, and the result information of history sample is fused to sample to be predicted Sample characteristics in, provide prediction to being suitably directed to sample to be predicted with reference to both history sample and sample to be predicted As a result.

Description of the drawings

From the detailed description below in conjunction with the accompanying drawings to the embodiment of the present invention, these and or other aspects of the invention and Advantage will become clearer and be easier to understand, wherein：

Fig. 1 shows according to an exemplary embodiment of the present invention to provide the frame of the system of prediction result based on machine learning Figure；

Fig. 2 shows the systems for providing prediction result based on machine learning according to another exemplary embodiment of the present invention Block diagram；

Fig. 3 shows according to an exemplary embodiment of the present invention to provide the flow of the method for prediction result based on machine learning Figure；

Fig. 4 shows the flow chart of the method for trained prediction model according to an exemplary embodiment of the present invention；

Fig. 5 shows the method for providing prediction result based on machine learning according to another exemplary embodiment of the present invention Flow chart；And

Fig. 6 shows the flow chart of the method for the training prediction model according to another exemplary embodiment of the present invention.

Specific implementation mode

In order to make those skilled in the art more fully understand the present invention, with reference to the accompanying drawings and detailed description to this hair Bright exemplary embodiment is described in further detail.

In an exemplary embodiment of the present invention, it is directed to sample to be predicted in the following manner and executes prediction：In addition to waiting for Except the attribute information for predicting sample itself, the attribute information of the history sample occurred before sample to be predicted is also additionally obtained And result information (for example, result information is obtained by predicting means) of the history sample about forecasting problem, by above-mentioned category Property information and result information are fused in the forecast sample of sample to be predicted so that machine learning model can be effectively with reference to going through Both history sample and sample to be predicted provide corresponding prediction result.

Machine learning is the inevitable outcome that artificial intelligence study develops to certain phase, is dedicated to the hand by calculating Section, improves the performance of system itself using experience.In computer systems, " experience " exists usually in the form of " data ", leads to Machine learning algorithm is crossed, " model " can be generated from data, that is to say, that empirical data is supplied to machine learning algorithm, just Model can be generated based on these empirical datas, when in face of new sample, model can provide corresponding judgement, that is, prediction result. It should be noted that exemplary embodiment of the present invention is to specific machine learning algorithm and without any restrictions.In addition, should also note Meaning, during training or application machine learning model, also using statistic algorithm, business rule and/or expertise etc., To further increase the accuracy of prediction result.

Fig. 1 shows according to an exemplary embodiment of the present invention to provide the frame of the system of prediction result based on machine learning Figure.Particularly, the forecasting system propose it is a kind of based on the information of sample to be predicted and its history sample come be directed to wait for it is pre- Test sample example execute prediction system for handling, in the system for handling, history sample about forecasting problem result information also by It is fused to the forecast sample of sample to be predicted, wherein the result information of certain history samples can be obtained by the means of prediction. System shown in FIG. 1 can be realized all by computer program with software mode, also can be realized by special hardware device, It can also be realized by way of software and hardware combining.Correspondingly, each device for forming system shown in FIG. 1 can be only relied on Computer program realizes the virtual module of corresponding function, can also be realized by hardware configuration the function general or Dedicated devices can also be that operation has the processor etc. of corresponding computer program.Using the system, history sample can be combined Result information utilize the attribute information of history sample, to effectively with reference to both history sample and sample to be predicted come compared with The prediction result of current sample to be predicted is adequately provided.

As described above, in an exemplary embodiment of the present invention, it, can when being predicted for current sample to be predicted It is judged in conjunction with both history sample and current sample, it particularly, can be under the guidance of the result information of history sample To consider the correlation attribute information of the history sample, enabling under limited hardware environment and data resource, utilize machine Device learning ways effectively hold the rule that history sample is embodied, and then provide accurate prediction result.

As shown in Figure 1, attribute information acquisition device 100 is for obtaining the attribute information of sample to be predicted and to be predicted The attribute information of the history sample occurred before sample.

Exemplary embodiment according to the present invention will utilize machine learning techniques, in conjunction with sample to be predicted (for example, current Sample) provide the prediction result of sample to be predicted with the relevant informations of one or more history samples.Sample mentioned here Refer to can (that is, about forecasting problem) is estimated in some respect object and/or affairs, correspondingly, sample may include but Be not limited to can be directed to whether cheat the transaction (for example, the transaction of credit card trade, deposit card, e-payment etc.) estimated, It can be directed to whether user execute displaying content that specific behavior estimated (for example, the advertisement that can click of user, user are commercially available The commodity etc. bought), the object (for example, physical signs etc.) or the like estimated of numberical range can be directed to.As showing Example, when sample indicates transaction sample (for example, buying behaviors such as the transaction of credit card trade, deposit card, e-payment), attribute letter Breath acquisition device 100 can obtain and merchandise every time related attribute information, for example, handing over incident position, the amount of money, trade company, quotient The information such as product.For example, in the case where whether it is fraudulent trading that forecasting problem is relationship trading, sample to be predicted can correspond to work as Preceding transaction, history sample can correspond to the previous transaction of the predetermined quantity occurred before current transaction and/or merchandise currently The previous transaction occurred in predetermined amount of time before.As described above, history sample refers to occurring before sample to be predicted Sample, as an example, history sample may include the one or more samples tightly occurred before sample to be predicted, particularly, Attribute information acquisition device 100 can calculating forward the day of trade from current transaction to be predicted, obtain the predetermined quantity occurred recently Historical trading and/or the historical trading occurred within nearest one section of predetermined time attribute information.In addition, history sample also may be used To be the one or more samples occurred before current sample to be predicted chosen according to other rules.

As an example, attribute information acquisition device 100 can obtain the data record of sample to be predicted and at least one history The data record of sample, these data records may include the category of each attribute about corresponding sample to be predicted or history sample Property information.

For example, above-mentioned data can be the data for prestoring or generating, can also be the data received from outside.These Data can relate to the identity information of object, for example, about information such as the identity of personnel, educational background, occupation, assets, contact methods.Or Person, these data can also refer to the relevant information of affairs, for example, about the turnover of deal contract, both parties, subject matter, The information such as loco.The content of the above attribute information is only as the example for explaining, in fact, being directed to sample to be predicted The specific sample (for example, transaction, state, network behavior etc.) of example and history sample, can obtain the attribute information of corresponding contents. That is the attribute information mentioned in exemplary embodiment of the present invention can relate to any sample (for example, any object or thing Business) performance in terms of certain or property, and be not limited to limit individual, object, tissue, unit, mechanism, project, event etc. Fixed or description.In fact, any can be by predicting that the data of attribute information of relevant issues can be applied using it as foundation In exemplary embodiment of the present invention.

In fact, attribute information acquisition device 100 can obtain structuring or the unstructured properties data of separate sources, example Such as, text data or numeric data etc..According to the specific sample of required prediction, attribute data may include deriving from various departments Data, for example, from business entity data, from bank and other financial mechanism data, derive from metadata provider Data, from internet (for example, social network sites) data, from mobile operator data, from APP run The data of quotient, the data from express company, data from credit institution etc..These data can pass through input unit It is input to attribute information acquisition device 100, or is automatically generated according to existing data by attribute information acquisition device 100, Or it can (for example, storage medium (for example, data warehouse) on network) obtains from network by attribute information acquisition device 100 , it is obtained from external data source in addition, the intermediate data switch of such as server can help to attribute information acquisition device 100 Take corresponding data.Here, the data of acquisition can be by data conversions such as text analysis models in attribute information acquisition device 100 Module is converted to the format being easily processed.It should be noted that attribute information acquisition device 100 can be configured as by software, hardware and/or The modules of firmware composition, these moulds certain module in the block or whole modules can be integrated into one or common cooperation with complete At specific function.

Result information acquisition device 200 is for obtaining result information of the history sample about forecasting problem, wherein needle To not having the history sample of the legitimate reading about forecasting problem, result information acquisition device 200 among the history sample Using the prediction result of history sample as the result information of history sample.

Here, for each history sample, result information acquisition device 200 can further obtain the history sample about The result information of forecasting problem.Wherein, result information is used to indicate the corresponding conclusion that history sample is directed to forecasting problem, the conclusion Identical or relevant content is may indicate that with the prediction result of sample to be predicted.For example, being related to whether transaction is to take advantage of in forecasting problem In the case of swindleness transaction, result information may be used to indicate whether each historical trading is fraudulent trading, and correspondingly, result information obtains Take device 200 that any mode appropriate can be taken to obtain the result information of each historical trading.As an example, result information obtains Take device 200 that can infer the result information of history sample according to the statistical result of a large amount of samples, for example, by significantly larger probability The result that can occur as history sample result information (for example, since arm's length dealing is apparent relative to the probability of fraudulent trading Therefore the result information of history sample can be set in advance as corresponding to arm's length dealing by higher)；As another example, as a result believe Ceasing acquisition device 200 can be using the legitimate reading of reflecting history sample actual conditions as the result information of corresponding history sample；Make For another example, result information acquisition device 200 can be according to other judgment criterions such as Expert Rules, the attribute based on history sample The result information of history sample is correspondingly arranged in information.It should be noted that above-mentioned example is not intended to limit the exemplary implementation of the present invention Any mode appropriate can be used to obtain the result information of history sample in the range of example, those skilled in the art.

Since when executing prediction for sample to be predicted, the legitimate reading for the history sample being based on may still can not , therefore, exemplary embodiment according to the present invention can be by such history for the history sample without legitimate reading The prediction result of sample is as its result information.For example, in the example that credit card steals brush, in order to predict that currently this brush card is No to steal brush, attribute information acquisition device 100 can obtain the current attribute information merchandised and previously merchandised, however, previously handing over Whether it is easily to steal brush to be most likely in still undetermined state (for example, user, which is also unaware that, has occurred robber's brush or user also Have not enough time to feedback steal brush generation), for this purpose, result information acquisition device 200 can will about this part previously transaction whether be The prediction result of brush is stolen as corresponding result information.As another example, recommendation items are directed to (for example, recommending in prediction user Commodity) click probability when, attribute information acquisition device 100 can obtain the case where first each for the previous period recommendation items are clicked, However, whether some samples can be marked and be clicked in these history samples for obtaining, but still due to nearest data It is untreated the reasons such as to finish, there can be the history sample that at least part not confirm click condition still, for this partial history sample Example, similarly, result information acquisition device 200 is using the prediction result of these history samples as their result information.

As can be seen that in an exemplary embodiment of the present invention, other than obtaining the attribute information of history sample, going back volume Particularly for the still uncertain history sample of result, corresponding prediction result is made for the outer result information for obtaining history sample For the result information of this partial history sample.By the above-mentioned means, can be under the guidance of the result information of relevant historical sample The attribute information or its statistical data for effectively applying each history sample, contribute to study/prediction of hoisting machine learning model Effect.

As described above, other than the history sample without legitimate reading, result information acquisition device 200 can be according to appointing What mode appropriate obtains the result information of other history samples.As an example, for the history sample with legitimate reading, Result information acquisition device 200 can be using the legitimate reading of history sample as the result information of history sample.Particularly, according to Exemplary embodiment of the present invention, in order to which the history sample for executing prediction for sample to be predicted and referring to can be greatly classified into two Kind, it is another one is the history sample (for example, the history sample occurred in the period earlier) for being labelled with legitimate reading Kind is the history sample (for example, the history sample occurred recently) that there is no method to obtain legitimate reading, is gone through accordingly for the first History sample, result information acquisition device 200 can obtain their legitimate reading information as a result, and as second of history sample Example, result information acquisition device 200 can obtain their prediction result information as a result.In this way, arbitrary history Sample can be used in executing prediction for sample to be predicted, this is particularly useful for the still unlabelled sample such as recent sample The prediction scene to play an important role.

Sample generating means 300 are for the attribute information of the sample to be predicted based on acquisition, the attribute of the history sample The result information of information and the history sample generates the forecast sample of sample to be predicted.Here, forecast sample is machine The basis that learning model is used to be predicted for sample to be predicted, that is to say, that correspond to sample to be predicted, sample generates dress 300 forecast samples for being made of generation multiple features are set, the feature can describe the spy of sample to be predicted from different perspectives Property (characteristic for including the history sample of sample to be predicted), for example, the feature can be the attribute information of each sample itself, Can also be the information field by these attribute informations obtain after characteristic processing or statistical disposition, it is particularly, described Feature may include the result information of history sample itself, may also comprise the result information of history sample and the knot of correlation attribute information Item is closed, for example, the end value for being counted and being generated to attribute information based on result information.

Exemplary embodiment according to the present invention, the feature of forecast sample will reflect the finger of the result information of history sample Lead effect, it should be appreciated that the concrete mode that result information participates in the Feature Engineering of forecast sample is unrestricted, and following example is only used for Play the role of explanation：

For example, sample generating means 300 can filter out at least part history sample according to the result information of history sample Example, and the attribute information next life of the attribute information based at least part history sample filtered out and sample to be predicted At the feature of the forecast sample of sample to be predicted.In this case, it can only consider that result meets the history sample of specified conditions, To improve the effect of model pre-estimating.For example, in the example of transaction swindling, can be possible for just according to result information to choose The historical trading often merchandised, the relativity between being merchandised by this partial history and currently being merchandised can effectively judge current Whether transaction is fraudulent trading.

In another example sample generating means 300 can be according to the result information of history sample to the corresponding category of the history sample Property information is weighted, and is generated based on the attribute information of the attribute information of the history sample after weighting and sample to be predicted The feature of the forecast sample of sample to be predicted.In this case, it is contemplated that the result information of each history sample effectively obtains The overall condition of history sample is taken, and is had together with the sample to be predicted attribute information of itself based on the overall condition of history sample Effect ground executes prediction.

In another example sample generating means 300 can be based respectively on the attribute letter of the attribute information of sample to be predicted, history sample Breath, the history sample result information generate the feature of the forecast sample of sample to be predicted.In this case, history sample The result information of example extends the dimension of feature space so that model can accordingly learn to more structurally sound knowledge, Jin Erti High prediction effect.

Above example is only not intended in any way to limit exemplary embodiment of the present invention as explaining, in fact, sample generates Device 300 can be directed to attribute information, the attribute information of relevant historical sample and the knot of the history sample of sample to be predicted Fruit information executes Feature Engineering to generate each feature of forecast sample, and here, those skilled in the art can be suitble to according to any Feature Engineering mode design the feature of forecast sample, such as, it is contemplated that service logic, the machine learning model of forecasting problem The various factors such as algorithm characteristic execute Feature Engineering, for example, can be to the attribute information of sample to be predicted or history sample Value carries out the processing of the various general characteristics engineerings such as combination, discretization, extraction part field value, rounding or statistics.It answers Note that exemplary embodiment of the present invention is not only restricted to any specific Feature Engineering scheme.

Furthermore, it is to be understood that exemplary embodiment according to the present invention, the feature of forecast sample can not only be based on sample to be predicted It is related can be also additionally based on other for attribute information, the attribute information of history sample and the result information of history sample of example Information.For example, result information acquisition device 200 can also obtain the confidence level of the result information of history sample, and correspondingly, sample life At device 300 can be based on the sample to be predicted of acquisition attribute information, the history sample attribute information, the history sample Result information and the confidence level of result information of the history sample generate the forecast sample of sample to be predicted.

Here, the confidence level of result information is alternatively referred to as reliability, confidence level, confidence coefficient.Make to population parameter When going out to estimate, due to the randomness of sample, conclusion is always uncertain, therefore, interval estimation is indicated using confidence level Assurance degree, here, the span of confidence interval is the positive function of confidence level, that is, the assurance degree required is bigger, certainly will obtain As soon as wider confidence interval, this correspondingly reduces the order of accuarcy of estimation.

Here, those skilled in the art can obtain the confidence of the result information of history sample in any suitable fashion Degree, for example, for the history sample without legitimate reading, result information acquisition device 200 can be based on the algorithm of prediction model Or the confidence level of the prediction result of history sample is obtained independently of the algorithm of prediction model, result information as history sample Confidence level；In addition, for the history sample with legitimate reading, result information acquisition device 200 can be by the true of history sample The confidence level of real result is set as indicating the preset value of high confidence level, the confidence level of the result information as history sample.

Correspondingly, sample generating means 300 can be using the confidence level of the result information of history sample as individual pre- test sample Eigen to extend the feature space of forecast sample, alternatively, sample generating means 300 can also by confidence level according to result information Similar mode is applied to Feature Engineering；Alternatively, sample generating means 300 can also be answered according to the mode different from result information Use confidence level.It should be noted that exemplary embodiment of the present invention has been not limited to form the feature of forecast sample and answer confidence level Concrete mode for Feature Engineering.

As an example, sample generating means 300 can generate sample to be predicted at least one of in the following manner The feature of forecast sample：

First way：Sample generating means 300 can be according to the history sample result information and the history sample The confidence level of the result information of example is gone through to filter out at least part history sample based on the described at least part filtered out The attribute information of the attribute information of history sample and sample to be predicted generates the feature of the forecast sample of sample to be predicted.At this In the case of kind, it can only consider that result meets the history sample of specified conditions, to improve the effect of model pre-estimating.For example, handing over In the example easily cheated, the historical trading for being possible for arm's length dealing can be chosen according to result information and its confidence level, passed through The relativity that this partial history is merchandised between current transaction can effectively judge whether currently merchandise is fraudulent trading.

The second way：According to the confidence of the result information of the history sample and the result information of the history sample Degree is weighted the respective attributes information of the history sample, and the attribute information based on the history sample after weighting and waits for The attribute information of sample is predicted to generate the feature of the forecast sample of sample to be predicted.In this case, it is contemplated that each history The result information of sample and its confidence level effectively obtain the overall condition of history sample, and the whole feelings based on history sample Condition efficiently performs prediction together with the sample to be predicted attribute information of itself.

The third mode：It is based respectively on the attribute information of sample to be predicted, the attribute information of the history sample, described goes through The confidence level of the result information of the result information of history sample and the history sample generates the forecast sample of sample to be predicted Feature.In this case, the result information of history sample and confidence level extend the dimension of feature space so that model energy More structurally sound knowledge is arrived in enough corresponding study, and then improves prediction effect.

Prediction result is provided device 400 and is used for using the prediction model trained based on machine learning techniques, pre- for waiting for The forecast sample of test sample example provides the prediction result of sample to be predicted.

That is, prediction result, which provides device 400, can be used for obtaining the prediction model by being trained based on machine learning The prediction result provided for sample to be predicted.Here, for the forecast sample of each sample to be predicted, prediction model can The prediction result of the sample to be predicted about forecasting problem is provided.Correspondingly, prediction result provides device 400 and can obtain by predicting The prediction result that model provides.Here, prediction result provides device 400 and can control prediction model to be directed to sample execution to be predicted Predict and thus to obtain the prediction result of prediction model, alternatively, prediction result device 400 is provided can be from positioned at external prediction mould Type receives corresponding prediction result.Prediction result, which provides device 400, to be supplied to user by the prediction result of acquisition, alternatively, in advance The prediction result of acquisition can be supplied to storage device, decision making device or other devices with into traveling by surveying result and providing device 400 The processing of one step.

Here, prediction model is the machine learning model trained based on training dataset, here, training dataset packet Include a large amount of training samples, the feature of each training sample constitutes consistent with forecast sample and includes current sample about forecasting problem Legitimate reading (that is, label).Citing is got on very well, it is assumed that the targeted forecasting problem of prediction model is whether current transaction is fraud Transaction can then utilize true sale whether a large amount of labeled fraud to form the training sample of prediction model, wherein very The fraudulent trading that occurred in fact can correspond to the positive sample of prediction model, and the non-fraudulent trading really occurred can correspond to it is pre- Survey the negative sample of model, as an example, the label of positive sample can value be 1, the label of negative sample can value be 0.About training The characteristic of sample needs to build according to consistent mode is corresponded to completely with forecast sample, particularly, each trains sample Originally current sample is can correspond to, correspondingly, attribute information that can be based on current sample, the history sample occurred before current sample The attribute information of example, the history sample about forecasting problem result information (or further also based on the history sample The confidence level of result information) generate each feature of training sample so that the feature of training sample and the feature of forecast sample It is upper completely the same (that is, uniform in the selection mode of history sample and the various aspects such as concrete mode of Feature Engineering constituting It causes).

As described above, in order to generate the feature of training sample, the result information for obtaining corresponding historical sample is needed, as The acquisition modes of preferred embodiment, the result information involved by training sample can correspond to the result information involved by forecast sample Acquisition modes.That is, can be tied by prediction result, legitimate reading, statistics according to consistent mode corresponding with forecast sample Fruit etc. is arranged in correspondence with the result information of history sample.

As an example it is assumed that executed for sample to be predicted estimate scene (for online Prediction user clicking rate or The scenes such as fraudulent trading) in, in order to generate the forecast sample of prediction model, not only need to obtain sample to be predicted and its history sample The attribute information of example, it is also necessary to obtain the result information of each history sample.However, for various reasons, it is likely that executing The legitimate reading information of all history samples can not be also obtained when estimating (for example, also having little time to collect or confirm when online Prediction The legitimate reading of all history samples), that is to say, that among the history sample that forecast sample is based on, some history Labeled legitimate reading, also at least part history sample do not have legitimate reading still to sample.Therefore, labeled true As a result history sample can be by legitimate reading information as a result, and the result information of other history samples may be from this The prediction result of a little history samples, as an example, can be the prediction result provided by prediction model.

In that case it is preferable that the training process and prediction process of prediction model can be consistently designed, in particular, Result information involved in training sample and forecast sample can be built according to consistent mode, so that training sample is on the scene Forecast sample is more approached on scape.For this purpose, although when building training sample, current sample and its history sample have had very It is real as a result, still, it is right in training sample for the history sample for using prediction result information as a result in forecast sample The history sample answered equally will be used as its result information using prediction result.For example, in stealing the example of brush about credit card, Assuming that forecast sample be based on be it is to be predicted transaction and nearest three months occur historical trading, wherein this month occur Historical trading does not have legitimate reading, needs to be used as result information using prediction result, and the historical trading of the other two moon Legitimate reading information as a result can be used.Correspondingly, when forming training sample, for the of that month history occurred of currently merchandising Transaction, also will be used as result information using prediction result, and legitimate reading conduct can be used in the historical trading of the other two moon Result information.

In the examples described above, the prediction result of history sample can be provided by prediction model, and here, prediction model can be used The mode of iteration is trained, the model currently trained obtained from continuous iteration, sustainable renewal training sample The prediction result of involved history sample.Correspondingly, it when the prediction model completed in application training is estimated, can be used Prediction model generates the prediction result of the history sample involved by forecast sample；Here, alternately, can also be used One takes turns model caused by iteration to provide the prediction result of the history sample involved by forecast sample.

That is, the prediction knot for the history sample that the forecast sample obtained by result information acquisition device 200 is based on Fruit can be provided by prediction model, alternatively, also can be by previous model corresponding with the last round of iteration of the prediction model It provides.

It, can be according to the side consistent with forecast sample in the training process of the prediction model accordingly, as example Formula, the legitimate reading of the history sample that training sample is based on or prediction result as the result information of the history sample, Wherein, the prediction result for the history sample that the training sample is based on is provided by the prediction model currently trained.Into one Step ground, in the training process, the prediction model is iterated training for training sample so that training sample is based on The prediction result of history sample constantly updated with iteration.

Prediction result, which provides device 400, can be supplied to the forecast sample of sample to be predicted the prediction mould as above trained Type, to obtain corresponding prediction result.As an example, system shown in FIG. 1 can provide sample to be predicted about prediction online The prediction result of problem.Correspondingly, attribute information acquisition device 100, result information acquisition device 200, sample generating means 300 There is provided device 400 with prediction result can online processing data.However, it should be understood that exemplary embodiment of the present invention is not limited In, whole system or in which certain devices can also be operated under off-line state.

As an example, the training process of above-mentioned prediction model also may be incorporated into prediction system according to an exemplary embodiment of the present invention In system.

Fig. 2 shows the systems for providing prediction result based on machine learning according to another exemplary embodiment of the present invention Block diagram.With reference to Fig. 2, the system comprises attribute information acquisition device 100, result information acquisition device 200, sample generating means 300, prediction result provides device 400 and model training apparatus 500.

As can be seen that compared to Figure 1, Fig. 2 further comprises model training apparatus 500, in addition to this, attribute information obtains dress Setting 100, result information acquisition device 200 and sample generating means 300 also will additionally execute the behaviour trained about prediction model Make.

Since attribute information acquisition device 100, result information acquisition device 200, sample being described in detail referring to Fig.1 Generating means 300 and prediction result are provided device 400 and are being executed operation when estimating using the complete prediction model of training, here will No longer aforesaid operations are repeated, only describe processing related with the training stage of prediction model below.

In order to complete the training of prediction model, a large amount of training samples need to be built based on true sample, wherein as showing Example, each training sample may indicate that a current legitimate reading of the sample about forecasting problem.The characteristic of training sample can The result information of attribute information, the history sample based on current sample and its history sample (or further considers the knot The confidence level of fruit information), and the label of training sample may indicate that the legitimate reading of current sample.

Particularly, attribute information acquisition device 100 can obtain the attribute information of current sample corresponding with training sample And the attribute information of the history sample occurred before current sample, here, attribute information acquisition device 100 also obtains currently Legitimate reading (that is, flag attribute) of the sample about forecasting problem.

Result information acquisition device 200 can obtain the history sample occurred before current sample about forecasting problem Result information, here, it is preferred that, result information acquisition device 200 can be corresponding according to the constituted mode of forecast sample Ground obtains the legitimate reading of above-mentioned history sample or prediction result is used as the result informations of these history samples.In addition to this, Alternately, result information acquisition device 200 can also accordingly obtain each result according to the acquisition modes of result information The confidence level of information.

It is the legitimate reading of attribute information, current sample that sample generating means 300 can be based on the current sample of acquisition, current The result information of the attribute information of the history sample of sample and the history sample (or is based further on the confidence of result information Degree) generate the training sample of current sample.It should be understood that sample generating means 300 can be passed through according to the characteristic Design of forecast sample It is handled by same Feature Engineering to generate the feature of training sample, in addition, sample generating means 300 can be by the true of current sample Label of the real result as training sample.

Model training apparatus 500 can be based on the training sample generated by sample generating means 300, according to scheduled machine Learning algorithm trains prediction model.Here, it should be noted that exemplary embodiment of the present invention does not limit the specific calculation of prediction model Method.Particularly, the mode of iteration can be used to train prediction model in model training apparatus 500, during this, is used as going through The prediction result (or together with its confidence level) of the result information of history sample is thus continually updated so that the characteristic quilt of training sample It constantly updates.For this purpose, in the model training of each round iteration, result information acquisition device 200 can be utilized and currently be trained The model model of last round of iteration (that is, corresponding to) obtain the prediction result of relevant historical sample, using as the history The result information of sample.For first round iteration, result information acquisition device 200 can obtain relevant historical otherwise The prediction result of sample, using the initial results information as the history sample.For example, result information acquisition device 200 can base The prediction result of relevant historical sample is set in the result statistical information of a large amount of samples, alternatively, result information acquisition device 200 The prediction result of relevant historical sample can be inferred to according to predetermined artificial rule, alternatively, result information acquisition device 200 can be random The prediction result of relevant historical sample is set.For such prediction result, result information acquisition device 200 will can accordingly be set Reliability uses as default or calculates corresponding confidence level in a predetermined manner.In addition, result information acquisition device 200 It can give up in first round iteration with the history sample of prediction result information as a result in training sample, and only with true As a result the history sample of information as a result.After first round iteration, you can needed using the prediction model trained to be directed to The history sample of prediction result is wanted to be predicted.As described above, model training apparatus 500 can constantly iteration prediction model, directly To meeting the corresponding condition of convergence, the prediction model so trained can learn to the history sample of various results under how needle Current sample is predicted.Training details about prediction model is said before can refer to about what prediction model itself carried out Bright, details are not described herein.

Correspondingly, the prediction model that training is completed can be supplied to prediction result to provide device by model training apparatus 500 400 so that prediction result, which provides device 400, to provide the pre- of sample to be predicted using the prediction model to be directed to forecast sample Survey result.

It as an example, in system shown in fig. 1 or fig. 2, may also include feedback device (not shown), waited for for reception pre- Legitimate reading of the test sample example about forecasting problem, wherein the legitimate reading be used to train together with corresponding sample to be predicted The prediction model.Particularly, in system such as shown in FIG. 1, sample to be predicted that feedback device can be received it Legitimate reading about forecasting problem is stored, and the legitimate reading of storage is supplied to the external trainer of prediction model Device, with re -training and/or update prediction model.Alternatively, in system such as shown in Fig. 2, feedback device can be by its institute Receive sample to be predicted be supplied to attribute information acquisition device 100 about the legitimate reading of forecasting problem, using as with wait for it is pre- The authentic signature of the corresponding training sample of test sample example.

It describes according to an exemplary embodiment of the present invention to provide prediction result based on machine learning hereinafter with reference to Fig. 3 Method flow chart.Here, as an example, the method can provide prediction knot of the sample to be predicted about forecasting problem online Fruit, correspondingly, at least part step need to execute online.

Method shown in Fig. 3 can be as shown in Figure 1 forecasting system execute, also can be completely by computer program with software Mode is realized or is realized by being stored with the computer-readable medium of the computer program.In addition, can also be matched by specific The computing device set executes method shown in Fig. 3.Since the processing for describing correlation technique step referring to Fig.1 above is thin Section, repeats, it should be appreciated that according to an exemplary embodiment of the present invention pre- for that will not refer again to Fig. 3 below the content of this part Survey method can equally cover described all processing details referring to Fig.1.

With reference to Fig. 3, in step S100, obtains the attribute information of sample to be predicted and occur before sample to be predicted The attribute information of history sample.In order to make it easy to understand, stealing the specific example of brush to be described, so below with reference to credit card And, it should be appreciated that exemplary embodiment of the present invention is not limited to any specific forecasting problem or related sample.Specifically Come, it is assumed that current this brush card transaction of sample instruction to be predicted, forecasting problem instruction are about current this brush card transaction No is the fraudulent trading for such as stealing brush, and history sample refers to the transaction of swiping the card occurred before current this brush card transaction.

As an example, credit card can be obtained online when the secondary generated attribute information currently merchandised of swiping the card, for example, swiping the card The amount of money, place of swiping the card, commodity of swiping the card, merchant identification of swiping the card etc.；In addition, can also obtain caused by the credit card swipes the card in the past The attribute information of historical trading, for example, can obtain nearest trimestral transaction attribute information or nearest N (N is whole more than 1 Number) transaction attribute information.

Next, in step s 200, obtaining result information of the history sample about forecasting problem, wherein for described The history sample for not having the legitimate reading about forecasting problem among history sample, using the prediction result of history sample as going through The result information of history sample.

Citing is got on very well, in this step, need to obtain about each historical trading whether be fraudulent trading result information, In, for the historical trading of fraudulent trading can not be confirmed whether it is in execution prediction fashion (that is, without true fraud result Historical trading), using the prediction result of the historical trading as its result information.Here, in all relevant historical tradings In the case of not having true fraud result, the result information that prediction result is used as historical trading can be used uniformly.As Example, prediction result here can be made by whether for predicting currently to merchandise, to be prediction model of fraudulent trading itself provide It, also can be corresponding to the last round of iteration by prediction model in the case where the prediction model is iterated training for optional mode Previous model provides the prediction result of historical trading.

However, in other cases, in all relevant historical tradings, it is also possible to which there are some with legitimate reading Historical trading merchandises for this partial history, the other information other than prediction result can be used and be used as result information, example Such as, conclusion will can really be cheated as corresponding result information.It should be noted that the result information of this partial history transaction is in addition to can Be historical trading legitimate reading except, can also be based on statistical probability obtained from great amount of samples, equally can also be The fraud possibility come out using model pre-estimating.

Next, in step S300, the attribute information of the sample to be predicted based on acquisition, the history sample attribute The result information of information and the history sample generates the forecast sample of sample to be predicted.

Citing is got on very well, and can correspond to each current transaction to generate corresponding forecast sample, feature is directed not only to currently The attribute information of transaction is directed to the attribute information of relevant historical transaction, in particular, when generating the feature of forecast sample, also Further combined with the fraud result information of each historical trading.

Here, the processing mode of any Feature Engineering appropriate can be used to generate the feature of forecast sample, as an example, The feature may include at least one among following item：

(1) at least one attributive character currently merchandised, for example, transaction amount, loco, tradable commodity, trader Family mark etc.；

(2) at least one attributive character of historical trading, for example, transaction amount, loco, tradable commodity, trader Family mark etc.；Here, can also be the statistics of attributes feature of historical trading, for example, average value/maximum value of transaction amount/most Small value etc.；

(3) result information of historical trading, for example, fraud result of the value between [0,1], wherein true fraud knot By can correspond to 1, true non-fraud conclusion can correspond to 0, and model prediction result or sample statistics result can be between 0 and 1 Probability value；Here, the feature can also be the result information statistical nature of historical trading；

(4) attributive character filtered out based on result information, for example, fraud end value is less than the historical trading of predetermined threshold Transaction amount；

(5) the statistics of attributes feature based on result information, for example, transaction amount is to cheat the value of result information as weight Weighted value.

It should be noted that the above project is only as an example, not a limit, for example, features described above producing method may also be combined with use, For example, the weighted feature after screening can further be obtained, for example, the historical trading for being less than predetermined threshold to fraud end value carries out It is handled using accordingly to cheat the value of result information as the weighting of weight to obtain corresponding feature.

With continued reference to Fig. 3, in step S400, the available prediction model trained based on machine learning techniques, for The forecast sample of sample to be predicted provides the prediction result of sample to be predicted.

As described above, as an example, prediction model here is for predicting whether currently merchandise is fraudulent trading.It will prediction Sample is supplied to the prediction model, can get the prediction result for the probability that the current transaction of instruction is fraudulent trading.

The flow of the method for trained prediction model according to an exemplary embodiment of the present invention is described hereinafter with reference to Fig. 4 Figure, here, a part of device in the system that method shown in Fig. 4 can be as shown in Figure 2 execute, and can also be instructed by individual model Practice device to execute, alternatively, the method can also be realized by computer program with software mode or by storing completely The computer-readable medium of computer program is stated to realize.In addition, also Fig. 4 institutes can be executed by the computing device of specific configuration The method shown.Similarly, due to describing the processing details of correlation technique step with reference to Fig. 2 above, for this part Fig. 4 will not be referred again to below content to repeat, it should be appreciated that model training method according to an exemplary embodiment of the present invention equally may be used Cover all processing details with reference to described in Fig. 2.

With reference to Fig. 4, in step S1000, obtains the attribute information of current sample and occur before current sample The attribute information of history sample.Here, current sample refers to the sample corresponding to current training sample.

Equally, for predicting the training process of prediction model of credit card robber's brush, in this step, some brush can be obtained The attribute information and the history that occurs before the transaction for blocking transaction are swiped the card the attribute information of transaction, for example, transaction amount, friendship Easy place, tradable commodity, transaction merchant identification etc..

In step S1100, authentic signature of the current sample about forecasting problem is obtained.

As an example, in this step, can obtain current transaction actually whether be fraudulent trading label as a result, its In, the label result of fraudulent trading can correspond to value 1, and the label result of non-fraudulent trading can correspond to value 0.

In step S2000, result information of the history sample about forecasting problem is obtained, here, as an example, can According to the mode consistent with forecast sample, the legitimate reading or prediction result for the history sample that training sample is based on are as institute State the result information of history sample, wherein the prediction result for the history sample that the training sample is based on can be by currently training The prediction model that goes out provides.

As an example, in this step, in training sample with the history in forecast sample with true fraud result Its true label can be used as its result information in transaction corresponding historical trading in sequential；And for training sample In with the historical trading for not having true fraud result in forecast sample in sequential corresponding historical trading, its prediction can be used As a result rather than its true fraud result is as its result information, prediction result here by the prediction model that currently trains Lai It provides.

In step S3000, attribute information based on the current sample and history sample that are obtained in step S1000, in step The label for the current sample that rapid S1100 is obtained generates current sample in the result information of the step S2000 history samples obtained Training sample.

Here, it should be noted that the feature of training sample and the feature of forecast sample are consistent, only in training sample also Additionally include the label information of current sample.

Next, in step S4000, using specific machine learning algorithm, prediction is trained based on training sample Model, here, exemplary embodiment according to the present invention can iteratively train prediction model, and correspondingly, step S4000 can be right Ying Yuyi takes turns iteration.

After a wheel training is completed, it is pre- can to judge whether the prediction model currently trained meets in step S5000 The fixed condition of convergence proceeds to step S6000 to export prediction model if meeting the condition of convergence.

If not meeting the condition of convergence, return to step S2000, using the prediction model currently trained come again The result information for obtaining history sample correspondingly updates each training sample in step S3000, is then held in step S4000 The model training of row next round.

Exemplary embodiment according to the present invention, in forecast sample/training sample of prediction model, in addition to combining history It, can also be further combined with the confidence level of result information, to be further ensured that the effect of model except the result information of sample.

It describes to provide prediction based on machine learning according to another exemplary embodiment of the present invention hereinafter with reference to Fig. 5 As a result the flow chart of method.As can be seen that the method for method shown in fig. 5 as shown in figure 3 is similar, the scene of the two application and The main body of execution can be same or like, and the step S100 in Fig. 5 can be identical as the step S100 in Fig. 3, the step in Fig. 5 S400 can be similar with the step S400 in Fig. 3, will no longer repeat here repetition or similar content.

Only emphasis describes the method for Fig. 5 and the distinct technology contents of method of Fig. 3 below.Particularly, in step In S210, other than obtaining history sample about the result information of forecasting problem, the result information of history sample is also obtained Confidence level.Here, as an example, for the history sample without legitimate reading, algorithm or independence that can be based on prediction model The confidence level that the prediction result of history sample is obtained in the algorithm of prediction model, the confidence of the result information as history sample Degree；For the history sample with legitimate reading, set the confidence level of the legitimate reading of history sample to indicate high confidence water Flat preset value, the confidence level of the result information as history sample.

Correspondingly, in step S310, the attribute information of the sample to be predicted based on acquisition, the history sample attribute The confidence level of the result information of information, the result information of the history sample and the history sample generates sample to be predicted Forecast sample.As an example, can according to result information similar mode, the confidence level of result information is also applied to together The Feature Engineering of forecast sample.

Here, again for predicting whether current transaction is fraudulent trading, in step S310, it is any appropriate to can be used The processing mode of Feature Engineering generate the feature of forecast sample, as an example, the feature may include among following item At least one of：

(3) result information and confidence level of historical trading, for example, the result information can be value between [0,1] Fraud result, wherein true fraud conclusion can correspond to 1, and true non-fraud conclusion can correspond to 0, model prediction result or Sample statistics result can be the probability value between 0 and 1；In addition, confidence level can weigh above-mentioned fraud result reliability Confidence value, true conclusion or the true non-fraud conclusion of cheating can have highest confidence value, in advance as legitimate reading Surveying the confidence level of result or statistical result can calculate according to corresponding method and be obtained；Here, the result of historical trading can also be used The statistical nature of information and/or confidence level；

(4) attributive character filtered out based on result information and/or confidence information, for example, fraud end value is less than The transaction amount of the historical trading of predetermined threshold, confidence level are tied higher than the transaction amount of the historical trading of predetermined threshold or fraud Fruit value is less than the transaction amount of predetermined threshold and confidence level higher than the historical trading of predetermined threshold；

(5) the statistics of attributes feature based on result information and/or confidence information, for example, transaction amount is to cheat result The value of information be the weighted value of weight, transaction amount using the value of confidence information as the weighted value of weight, transaction amount with Cheat the weighted value for being combined into weight of the value of result information and the value of confidence information.

It should be noted that the above project is only as an example, not a limit, for example, features described above producing method may also be combined with use.

Correspondingly, Fig. 6 shows the flow chart of the method for the training prediction model according to another exemplary embodiment of the present invention. As can be seen that method method as shown in fig. 4 shown in fig. 6 is similar, the scene of the two application and the main body of execution can be identical Or it is similar, step S1000, S1100, S4000, S5000 and S6000 in Fig. 6 can or phases identical as the corresponding steps in Fig. 4 Seemingly, no longer repetition or similar content will be repeated here.

Only emphasis describes the method for Fig. 6 and the distinct technology contents of method of Fig. 4 below.Particularly, in step In S2100, other than obtaining history sample about the result information of forecasting problem, the result information of history sample is also obtained Confidence level.Correspondingly, in step S3100, the attribute information of the current sample based on acquisition, the history sample attribute letter Breath, the result information of the history sample and the history sample the confidence level of result information generate the instruction of current sample Practice sample.It should be noted that the calculation of confidence level is consistent with the content described with reference to Fig. 5, and the Feature Engineering of training sample Can be consistent with the content with reference to Fig. 5 descriptions, it will no longer repeat here.

Above by reference to Fig. 1 to Fig. 6 describe it is according to an exemplary embodiment of the present invention based on machine learning come provide wait for it is pre- Test sample example is about the method and system of the prediction result of forecasting problem and corresponding model training method and system.It should be understood that The above method can be realized by the program being recorded in computer-readable media, correspondingly, exemplary implementation according to the present invention Example, it is possible to provide a kind of to provide computer-readable Jie of prediction result of the sample to be predicted about forecasting problem based on machine learning Matter, record is useful for executing the computer program of following methods step on the computer-readable medium：(A) it obtains to be predicted The attribute information of the attribute information of sample and the history sample occurred before sample to be predicted；(B) the history sample is obtained Result information of the example about forecasting problem, wherein for the true knot not having among the history sample about forecasting problem The history sample of fruit, using the prediction result of history sample as the result information of history sample；(C) sample to be predicted based on acquisition The result information of the attribute information of example, the attribute information of the history sample and the history sample generates sample to be predicted Forecast sample；(D) using the prediction model that is trained based on machine learning techniques, for sample to be predicted forecast sample come The prediction result of sample to be predicted is provided.

Computer program in above computer readable medium can be in client, host, agent apparatus, server etc. Run in the environment disposed in computer equipment, it should be noted that the computer program can be additionally used in execute in addition to above-mentioned steps with Outer additional step or executed when executing above-mentioned steps more specifically handles, these additional steps and is further processed Content is described referring to figs. 1 to Fig. 6, here in order to avoid repetition will be repeated no longer.

Correspondingly, the above-mentioned system for providing prediction result of the sample to be predicted about forecasting problem based on machine learning The operation of computer program can be completely dependent on to realize corresponding function, that is, the function structure of each device and computer program In each step it is corresponding so that whole system is called by special software package (for example, libraries lib), corresponding to realize Function.

On the other hand, Fig. 1 and each device shown in Fig. 2 and unshowned relevant apparatus can also be by hardware, soft Part, firmware, middleware, microcode or its arbitrary combination are realized.When being realized with software, firmware, middleware or microcode, use It can be stored in the computer-readable medium of such as storage medium, make in the program code or code segment for executing corresponding operating Corresponding operation can be executed by reading and running corresponding program code or code segment by obtaining processor.

Here, exemplary embodiment of the present invention is also implemented as computing device, which includes storage unit And processor, set of computer-executable instructions conjunction is stored in storage unit, when the set of computer-executable instructions is closed by institute When stating processor and executing, executes and above-mentioned provide the side of prediction result of the sample to be predicted about forecasting problem based on machine learning Method.

Particularly, the computing device can be deployed in server or client, can also be deployed in distributed network On node apparatus in network environment.In addition, the computing device can be PC computers, board device, personal digital assistant, intelligence Can mobile phone, web applications or other be able to carry out the device of above-metioned instruction set.

Here, the computing device is not necessarily single computing device, can also be it is any can be alone or in combination Execute the device of above-metioned instruction (or instruction set) or the aggregate of circuit.Computing device can also be integrated control system or system A part for manager, or can be configured as with Local or Remote (for example, via wireless transmission) with the portable of interface inter-link Formula electronic device.

In the computing device, processor may include central processing unit (CPU), graphics processor (GPU), may be programmed and patrol Collect device, dedicated processor systems, microcontroller or microprocessor.As an example, not a limit, processor may also include simulation Processor, digital processing unit, microprocessor, multi-core processor, processor array, network processing unit etc..

It is above-mentioned about providing sample to be predicted about institute in the method for the prediction result of forecasting problem based on machine learning Certain operations of description can realize that certain operations can be realized by hardware mode by software mode, in addition, can also pass through The mode of software and hardware combining realizes these operations.

Processor can run the instruction being stored in one of storage unit or code, wherein the storage unit can be with Store data.Instruction and data can be also sent and received via Network Interface Unit and by network, wherein the network connects Any of transport protocol can be used in mouth device.

Storage unit can be integral to the processor and be integrated, for example, RAM or flash memory are arranged in integrated circuit microprocessor etc. Within.In addition, storage unit may include independent device, such as, external dish driving, storage array or any Database Systems can Other storage devices used.Storage unit and processor can be coupled operationally, or can for example by the ports I/O, Network connection etc. communicates so that processor can read the file being stored in storage unit.

In addition, the computing device may also include video display (such as, liquid crystal display) and user's interactive interface is (all Such as, keyboard, mouse, touch input device etc.).The all components of computing device can be connected to each other via bus and/or network.

It is above-mentioned about being provided involved by method of the sample to be predicted about the prediction result of forecasting problem based on machine learning And operation can be described as it is various interconnection or coupling functional blocks or function diagram.However, these functional blocks or function diagram Single logic device can be equably integrated into or operated according to non-exact boundary.

Particularly, as described above, according to an exemplary embodiment of the present invention provide sample to be predicted based on machine learning Example may include storage unit and processor about the computing device of the prediction result of forecasting problem, and calculating is stored in storage unit Machine executable instruction set executes following step when set of computer-executable instructions conjunction is executed by the processor： (A) attribute information of sample to be predicted and the attribute information of the history sample occurred before sample to be predicted are obtained；(B) it obtains Take result information of the history sample about forecasting problem, wherein do not have about prediction among the history sample The history sample of the legitimate reading of problem, using the prediction result of history sample as the result information of history sample；(C) it is based on obtaining The result information next life of the attribute information of the sample to be predicted taken, the attribute information of the history sample and the history sample At the forecast sample of sample to be predicted；(D) using the prediction model trained based on machine learning techniques, for sample to be predicted Forecast sample the prediction result of sample to be predicted is provided.

It should be noted that have been combined above Fig. 1 to Fig. 6 describe it is according to an exemplary embodiment of the present invention be based on machine learning Details is managed everywhere in method of the sample to be predicted about the prediction result of forecasting problem to provide, will not be described in great detail calculating dress here Set processing details when executing each step.

Each exemplary embodiment of the present invention is described above, it should be appreciated that foregoing description is merely exemplary, not Exhaustive, and present invention is also not necessarily limited to disclosed each exemplary embodiments.Without departing from scope and spirit of the present invention Sample under, many modifications and changes will be apparent from for those skilled in the art.Therefore, originally The protection domain of invention should be subject to the scope of the claims.

Claims

1. a kind of method providing prediction result of the sample to be predicted about forecasting problem based on machine learning, including：

(A) attribute information of sample to be predicted and the attribute information of the history sample occurred before sample to be predicted are obtained；

(B) result information of the history sample about forecasting problem is obtained, wherein for not having among the history sample The history sample of legitimate reading about forecasting problem, using the prediction result of history sample as the result information of history sample；

(C) attribute information of the sample to be predicted based on acquisition, the attribute information of the history sample and the history sample Result information generate the forecast sample of sample to be predicted；And

(D) it using the prediction model trained based on machine learning techniques, is waited for for the forecast sample of sample to be predicted to provide Predict the prediction result of sample.

2. the method for claim 1, wherein in step (B), the confidence of the result information of history sample is also obtained Degree, also,

In step (C), the attribute information of the sample to be predicted based on acquisition, the history sample attribute information, described go through The confidence level of the result information of the result information of history sample and the history sample generates the forecast sample of sample to be predicted.

3. method as claimed in claim 2, wherein in step (B), the prediction result of the history sample is by the prediction Model or previous model corresponding with the last round of iteration of the prediction model provide.

4. method as claimed in claim 3, wherein in step (B), for the history sample with legitimate reading, will go through Result information of the legitimate reading of history sample as history sample.

5. method as claimed in claim 4, wherein the prediction model has following training process, in the training process In, according to the mode consistent with forecast sample, the legitimate reading or prediction result of the history sample that training sample is based on are made For the result information of the history sample, wherein the prediction result for the history sample that the training sample is based on is by currently instructing The prediction model practised provides.

6. method as claimed in claim 5, wherein in the training process, the prediction model for training sample into Row iteration is trained so that the prediction result for the history sample that training sample is based on is constantly updated with iteration.

7. method as claimed in claim 2, wherein generate the pre- of sample to be predicted at least one of in the following manner The feature of test sample sheet：

(C1) it is filtered out according to the confidence level of the result information of the history sample and the result information of the history sample At least part history sample, and the attribute information based at least part history sample filtered out and sample to be predicted The attribute information of example generates the feature of the forecast sample of sample to be predicted；

(C2) it is gone through to described according to the confidence level of the result information of the history sample and the result information of the history sample The respective attributes information of history sample is weighted, and the attribute information based on the history sample after weighting and sample to be predicted Attribute information generates the feature of the forecast sample of sample to be predicted；And

(C3) it is based respectively on the knot of the attribute information of sample to be predicted, the attribute information of the history sample, the history sample The confidence level of the result information of fruit information and the history sample generates the feature of the forecast sample of sample to be predicted.

8. a kind of system providing prediction result of the sample to be predicted about forecasting problem based on machine learning, including：

Attribute information acquisition device, attribute information for obtaining sample to be predicted and what is occurred before sample to be predicted go through The attribute information of history sample；

Result information acquisition device, for obtaining result information of the history sample about forecasting problem, wherein for described Do not have the history sample of the legitimate reading about forecasting problem among history sample, result information acquisition device is by history sample Result information of the prediction result as history sample；

Sample generating means, for the attribute information of the sample to be predicted based on acquisition, the history sample attribute information with And the result information of the history sample generates the forecast sample of sample to be predicted；And

Prediction result provides device, for utilizing the prediction model trained based on machine learning techniques, for sample to be predicted Forecast sample the prediction result of sample to be predicted is provided.

9. a kind of providing the computer-readable medium of prediction result of the sample to be predicted about forecasting problem based on machine learning, Wherein, record is useful for executing the computer program of following steps on the computer-readable medium：

10. a kind of providing the computing device of prediction result of the sample to be predicted about forecasting problem based on machine learning, including Storage unit and processor are stored with set of computer-executable instructions conjunction in storage unit, when the computer executable instructions When set is executed by the processor, following step is executed：