CN112785331B - Robust recommendation method and system for resisting injection attack by combining evaluation text - Google Patents
Robust recommendation method and system for resisting injection attack by combining evaluation text Download PDFInfo
- Publication number
- CN112785331B CN112785331B CN202110018824.1A CN202110018824A CN112785331B CN 112785331 B CN112785331 B CN 112785331B CN 202110018824 A CN202110018824 A CN 202110018824A CN 112785331 B CN112785331 B CN 112785331B
- Authority
- CN
- China
- Prior art keywords
- evaluation
- model
- scoring
- user
- personalized
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000011156 evaluation Methods 0.000 title claims abstract description 200
- 238000000034 method Methods 0.000 title claims abstract description 50
- 238000002347 injection Methods 0.000 title claims abstract description 36
- 239000007924 injection Substances 0.000 title claims abstract description 36
- 238000012549 training Methods 0.000 claims abstract description 43
- 238000001514 detection method Methods 0.000 claims abstract description 39
- 230000006870 function Effects 0.000 claims description 14
- 238000013528 artificial neural network Methods 0.000 claims description 8
- 230000002159 abnormal effect Effects 0.000 claims description 5
- 238000004590 computer program Methods 0.000 claims description 5
- 230000015654 memory Effects 0.000 claims description 5
- 239000000758 substrate Substances 0.000 claims 1
- XLYOFNOQVPJJNP-UHFFFAOYSA-N water Substances O XLYOFNOQVPJJNP-UHFFFAOYSA-N 0.000 abstract description 17
- 238000004422 calculation algorithm Methods 0.000 abstract description 6
- 238000003062 neural network model Methods 0.000 abstract description 2
- 238000001914 filtration Methods 0.000 description 10
- 239000013598 vector Substances 0.000 description 7
- 239000011159 matrix material Substances 0.000 description 6
- 230000006399 behavior Effects 0.000 description 5
- 230000001537 neural effect Effects 0.000 description 4
- 238000007637 random forest analysis Methods 0.000 description 4
- 238000011161 development Methods 0.000 description 3
- 238000005457 optimization Methods 0.000 description 3
- 238000013459 approach Methods 0.000 description 2
- 230000007547 defect Effects 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 230000003993 interaction Effects 0.000 description 2
- 230000002452 interceptive effect Effects 0.000 description 2
- 239000000243 solution Substances 0.000 description 2
- 230000000007 visual effect Effects 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000000295 complement effect Effects 0.000 description 1
- 238000013499 data model Methods 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 230000003287 optical effect Effects 0.000 description 1
- 238000012552 review Methods 0.000 description 1
- 238000005096 rolling process Methods 0.000 description 1
- 230000006403 short-term memory Effects 0.000 description 1
- 230000001360 synchronised effect Effects 0.000 description 1
- 238000012360 testing method Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q30/00—Commerce
- G06Q30/02—Marketing; Price estimation or determination; Fundraising
- G06Q30/0201—Market modelling; Market analysis; Collecting market data
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/90—Details of database functions independent of the retrieved data types
- G06F16/901—Indexing; Data structures therefor; Storage structures
- G06F16/9024—Graphs; Linked lists
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/90—Details of database functions independent of the retrieved data types
- G06F16/906—Clustering; Classification
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q30/00—Commerce
- G06Q30/06—Buying, selling or leasing transactions
- G06Q30/0601—Electronic shopping [e-shopping]
- G06Q30/0631—Item recommendations
Landscapes
- Engineering & Computer Science (AREA)
- Business, Economics & Management (AREA)
- Theoretical Computer Science (AREA)
- Accounting & Taxation (AREA)
- Finance (AREA)
- Development Economics (AREA)
- Databases & Information Systems (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Strategic Management (AREA)
- Data Mining & Analysis (AREA)
- General Engineering & Computer Science (AREA)
- Economics (AREA)
- Marketing (AREA)
- General Business, Economics & Management (AREA)
- Entrepreneurship & Innovation (AREA)
- Software Systems (AREA)
- Game Theory and Decision Science (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention designs a robust recommendation method and a system for resisting injection attack by combining evaluation texts. The method comprises the following steps: 1) Training an evaluation scoring prediction model by using the known information and the calculated statistical information on the graph by adopting a graph neural network model; 2) Training the personalized evaluation text generation model by using the text generation model and predictive scoring; 3) And training an attack detection model according to the errors of the predicted value and the true value of the evaluation scoring prediction model and the personalized evaluation text generation model. The invention integrates the evaluation scoring prediction model, the personalized evaluation text generation model and the attack detection model into a whole, so that three tasks are cooperated and mutually promoted. According to the scheme, evaluation text information is fully utilized, feedback data generated by 'water army' users are divided and utilized in finer granularity, meanwhile, the influence of 'injection attack' in an electronic commerce recommendation platform on recommendation accuracy can be automatically relieved, and the robustness of a recommendation algorithm is improved.
Description
Technical Field
The method belongs to the technical field of information. The method is mainly used for analyzing the interests and commodity characteristics of a user by combining the evaluation text contents aiming at a recommendation system in an e-commerce website, automatically identifying which evaluations come from injection attack, and reducing the influence of the evaluations in the recommendation system. The invention utilizes three parts of personalized text content prediction, evaluation scoring prediction and attack detection to mutually promote, and can automatically relieve the problem of attack injection in the electronic commerce recommendation system, thereby improving the robustness and accuracy of the recommendation system.
Background
With the development of the e-commerce industry, more and more users select online shopping. A large number of commodities are available for users to select on the electronic commerce platform, and in order to improve user experience and accelerate the purchase decision process of the users, the electronic commerce platform uses a recommendation system to guess the interests of the users and provides personalized commodity recommendation for the users. It is very necessary for the e-commerce platform to ensure the accuracy of the recommendation system.
Among the many recommended methods, collaborative filtering is a mainstream method, which can be further divided into memory-based collaborative filtering and model-based collaborative filtering. The user-based collaborative filtering and the commodity-based collaborative filtering are collectively referred to as memory-based collaborative filtering, and it is assumed that similar users have similar interests and similar commodities have similar characteristics, respectively. The model-based collaborative filtering is to represent the scoring of the commodity by the user by a matrix and decompose the scoring into the product of two low-dimensional matrices, and the essence is to carry out matrix complement on the scoring matrix. Recently, the development of deep learning has promoted the development of a recommendation method, and a series of deep collaborative filtering models improve the effect of the traditional recommendation method by jointly learning a user vector representation representing the user interests and a commodity vector representation representing the commodity characteristics. Users, commodities and feedback data of the users on the commodities in the electronic commerce platform naturally form a graph structure, the graph neural network is an effective method for modeling the graph structure, and a series of deep collaborative filtering methods based on the graph neural network further improve the recommendation effect.
However, the user feedback data in the e-commerce platform is not all of high quality and trusted, and some merchants may employ water armies to release false ratings, giving an unfair score, in order to promote or deface the target commodity, a behavior called "injection attack". Some studies have shown that the presence of "injection attacks" can reduce the accuracy of the recommendation system, thereby affecting market fairness.
Note that in this case, efforts have been made to mitigate the impact of "injection attacks" on the recommendation system. Some approaches focus on how to improve the recommendation system robustness by using a robust model; some methods first detect false feedback from the user, then remove the false feedback, and then run the recommendation system. A recent work (GraphRfi) indicates that the two methods respectively have the defects that the model is more trained and worse after the user auxiliary information is difficultly utilized and the normal data is misjudged as false and deleted, and the two methods are combined, so that the two methods can be taken together and promoted mutually. Specifically, graphRfi is divided into two tasks of evaluation scoring prediction and water army detection, and learning is performed by using a graph rolling network (Graph Convolutional Networks) and a neural random forest (Neural Random Forests) respectively, and the tasks are combined training. In the evaluation scoring task, if the evaluation scoring of a user is significantly different from the predicted value, the user is highly likely to be a water army user; in a water force detection task, if a user is classified as a water force, the confidence level of the user may be low, and the influence of the user in an evaluation scoring task needs to be reduced.
It is noted that the collaborative filtering method described above in connection with the graph neural network is still inadequate in terms of the utilization of the information on the graph. The method mainly utilizes the attribute information of nodes (users and commodities) and user scoring on edges (feedback behaviors of the users to the commodities) to predict evaluation scoring. However, feedback of the user to the commodity not only points the commodity, but also evaluates the commodity by the user, and contains rich text information. The rich text information can more accurately reflect the interests of the user and the characteristics of the commodity.
The key point of the patent is to expand the existing graph-based recommendation algorithm and improve the capturing capability of the recommendation algorithm on the evaluation text information.
Disclosure of Invention
The invention provides an automatic robust recommendation method and system aiming at injection attack in a recommendation system and combining rich evaluation text content. The method integrates personalized text content prediction, evaluation scoring prediction and attack detection, and can promote each other.
In the following description, the data model of the present invention is denoted as graph g= (U I, E), where U represents a set of user nodes, I represents a set of commodity nodes, and E represents a set of edges (a relationship that a user generates feedback on a commodity). The user node U epsilon U and the commodity node I epsilon I are provided with attribute information, and the side E epsilon E is provided with user feedback information (comprising evaluation text content C epsilon C and user scoring R epsilon R for commodities, wherein C is an evaluation text content set, R is a user scoring set for commodities) and a specific timestamp T epsilon T (T is a timestamp set) for generating feedback.
The technical scheme adopted by the invention is as follows:
the invention firstly provides a robust recommendation model training method for resisting injection attack by combining an evaluation text, which comprises the following steps:
training an evaluation scoring prediction model;
training a personalized evaluation text generation model according to the evaluation scoring predicted by the evaluation scoring prediction model and the generated user representation and commodity representation;
Training an attack detection model according to errors of a predicted value and a true value of the evaluation scoring prediction model and the personalized evaluation text generation model, and detecting injection evaluation;
and according to the injection evaluation detection result output by the attack detection model, reducing the influence weight of the corresponding evaluation in the evaluation scoring prediction model and the personalized evaluation text generation model.
Further, the training evaluation scoring prediction model is obtained by training the graphic neural network according to the user attribute, the commodity attribute, the evaluation content of commodities given by the user, the scoring and the time for generating feedback by the user, and obtaining the representation of the user, the commodities and the feedback edges; the user attribute is a statistic capable of reflecting whether the user is abnormal.
Further, the above steps are divided into three tasks: the method comprises a personalized evaluation text generation task, an evaluation scoring prediction task and an attack detection task, wherein the three tasks interact and promote each other.
Specifically, the personalized evaluation text generation task may assist in evaluating the scoring prediction task; the evaluation scoring prediction task can guide the personalized evaluation text generation task; if the predicted integrated value of the first two tasks is significantly different from the true value, the evaluation is quite likely to be injection evaluation; in the attack detection task, if one evaluation is classified as an injection evaluation, the influence of the evaluation in the first two tasks (i.e., the control influence weight) needs to be reduced.
Further, as the three tasks are mutually influenced and promoted, the invention designs a unified loss function, so that the three tasks are unified and end-to-end collaborative training is carried out.
Specifically, for the evaluation scoring prediction task, we train the evaluation scoring prediction model S (U, I, C, R, T). Fwdarw.R 'so that the predicted score R' of the edge to be tested is close to the true scoreAnd when training the evaluation scoring prediction model, the influence weight of false evaluation scoring needs to be reduced according to the output of the attack detection model.
Evaluating a loss function of the scoring predictive model:
Where p e represents the probability that edge e is true, r' e represents the predictive value of edge e generated by the scoring predictive model, Representing the true score of edge e.
For the personalized evaluation text generation task, we train a personalized evaluation text generation model cg (U, I, C, R, T) →C 'so that the generated evaluation C' of the edge to be tested is close to the real evaluationFurther, the predictive value r' e generated by the evaluation scoring predictive model is added for guidance when the personalized evaluation text generation model is trained. And when the personalized evaluation text generation model is trained, the influence weight of false evaluation content needs to be reduced according to the output of the attack detection model. Meanwhile, an auxiliary model f for predicting and evaluating the evaluation by using the evaluation content is introduced, the user representation and the commodity representation generated by using the evaluation scoring prediction model are personalized, and the evaluation scoring prediction model and the personalized evaluation text generation model are trained together by using the same set of user representation and commodity representation, so that the evaluation scoring prediction task and the personalized evaluation text generation task are mutually coordinated.
Loss function of personalized evaluation text generation model:
Where p e represents the probability that edge e is true, c' e represents the rating of edge e generated by the personalized rating text generation model, Representing a true evaluation of edge e.
For the attack detection task, according to the difference condition of the edge prediction scoring and the generated evaluation text content, the true scoring and the true text content in the first two tasks, updating the edge representation (thereby affecting the training of the graph neural network in the evaluation scoring prediction task and the training of the generation model in the personalized evaluation text generation task in the new updating). Meanwhile, based on the updated edge representation, the training attack detection model predicts whether the edge is from an attack (p e represents the probability that edge e is true).
Loss function (log loss) of attack detection model:
Where y e represents a label of whether the edge is real, e| represents the total number of edges, and p e represents the probability that the edge E is real.
The model overall loss function is composed of three tasks, and the three tasks are uniformly optimized in a training stage.
Specifically, the overall loss function:
L＝Lrating+α·Lcontent_generate+β·Lfraudster
wherein, alpha and beta represent super parameters for controlling three loss ratios.
Based on the training method, the invention provides a robust recommendation method for resisting injection attack combined with evaluation text, the method comprises the steps of predicting user-commodity to be predicted to score corresponding evaluation by using an injection attack resistant robust recommendation model combined with an evaluation text after training by the method, and recommending commodity according to the evaluation score.
Based on the same inventive concept, the invention also provides a robust recommendation system for resisting injection attack combined with an evaluation text, which comprises:
the model training module is used for training a robust recommendation model for resisting injection attack by combining the evaluation text by adopting the method;
And the recommendation module is used for predicting the corresponding evaluation scoring of the user-commodity to be predicted by using the robust recommendation model of the injection attack resistance combined with the evaluation text after the training is finished, and recommending the commodity according to the evaluation scoring.
The beneficial effects of the invention are as follows:
Aiming at the problem that the accuracy of a recommendation system is reduced due to 'injection attack', the invention provides a unified training method integrating evaluation scoring prediction task, personalized evaluation text generation task and attack detection task. The existing methods only pay attention to improving the robustness of the recommendation algorithm or apply the recommendation algorithm after attack is removed through a water army detection model, and the existing methods have the defects that the model is more trained and worse after user auxiliary information is difficultly utilized and normal data is misjudged as false and deleted. Recent work GraphRfi, while combining both the evaluation scoring prediction task and the water army detection task for unified training, has two problems that have not been solved: first, some users, while employed by some merchants to issue "injection attack" ratings, may still have normal transaction, rating behavior, all relevant feedback for users identified as "water forces" are de-weighted, losing their normal feedback data; second, in the utilization of feedback data, the model only utilizes scoring data of the commodity by the user, while in the electronic market, text content of the user's published evaluation of the commodity may reflect more information about the interests of the user and information about the characteristics of the commodity, which is not utilized in the previous model.
The technology focuses on improving the robustness of the recommendation algorithm. Compared with the prior method, the method fully utilizes the evaluation text information in the aspect of information source utilization; in the aspect of utilizing attack detection to reduce weight, the method fully considers the characteristic that some 'water army' users still have normal feedback data on non-target commodities, and compared with 'water army detection', the 'attack detection' is used for distinguishing the feedback data generated by the 'water army' users in finer granularity; in the aspect of model training, the problem of attack injection in the electronic commerce recommendation system can be automatically relieved by mutual promotion of three parts of evaluation scoring prediction, personalized evaluation text generation and attack detection, so that the robustness and the accuracy of the recommendation system are improved.
Drawings
Fig. 1: and an interactive frame diagram among the evaluation scoring prediction model, the personalized evaluation text generation model and the attack detection model.
Fig. 2: e-commerce feedback network schematic diagram.
Detailed Description
The invention is further described in more detail below based on an e-commerce recommendation platform.
The invention discloses a robust recommendation method for evaluating text content in combination with injection attack, which comprises the following steps:
1) Statistics are calculated as user attributes that can reflect whether the user is abnormal based on the scoring matrix of the merchandise and the time stamp that generated the scoring.
2) According to the user attribute and the commodity attribute, the user obtains the evaluation content and scoring of the commodity and the feedback time of the user, and the graphic neural network is used for training to obtain the representation of the user, the commodity and the feedback edge, so as to obtain an evaluation scoring prediction model; and further predicting the corresponding scoring (scoring matrix completion) of the user-commodity to be tested by using the evaluation scoring prediction model.
3) The personalized evaluation text generation model is utilized to generate personalized evaluation text (the generated evaluation content is related to user interests and commodity characteristics, specifically, the personalized evaluation text generation model is realized by adding user representations and commodity representations in an evaluation scoring prediction model into a coding part of the text generation model), and the evaluation scoring prediction result in the step 2) is added for guidance when the generation model is trained.
4) An auxiliary model f scored by evaluation content prediction evaluation is trained by using each evaluation content c and scoring r on the edge, so that f (c) and r are as close as possible. f (c) represents an evaluation score predicted by the evaluation content c. The auxiliary model f is used for optimizing user representation and commodity representation in the evaluation scoring prediction model through training of the personalized evaluation text generation model, so that the auxiliary evaluation scoring prediction model is used for evaluation scoring.
5) Comprehensively considering the difference between the generated results of the personalized evaluation text generation model and the evaluation scoring prediction model and the true value, performing attack detection tasks, namely performing false prediction on the edge (the larger the difference is, the more likely to be false edge, the influence weight of the edge in the first two models needs to be correspondingly reduced), generating a representation vector representing the difference, splicing the representation vector into the edge vector, thereby acting on the personalized evaluation text generation model and the evaluation scoring prediction model, finally obtaining the scoring corresponding to the user-commodity pair to be predicted through the trained evaluation scoring prediction model, and recommending the commodity according to the scoring.
A specific application example is provided below, comprising the steps of:
1) Fig. 2 shows an example of an e-commerce feedback network. U 1,u2,…,u8 in the figure represents a user node; i 1,i2,…,i6 represents a commodity node; the connection between the user node and the commodity node, the solid line represents the real feedback relation of the known user in the training set to the commodity, and the dotted line represents the preference degree of the user to be predicted in the testing set to the commodity; the edges are provided with time stamps for feedback relation occurrence; and c u,i,ru,i on the edge respectively represents the evaluation text of the commodity published by the user and the commodity is scored by the user.
2) And calculating some statistical information for the user as attributes reflecting whether the user is abnormal or not according to the feedback scoring and the time when the feedback behavior occurs. Such as the total number of items the user rated, the proportion of the user scored 1 to 5, the entropy of the user scoring, the proportion of all the user's ratings "praise", the time interval for the user to take place the feedback action, etc.
3) And training an evaluation scoring prediction model by utilizing abundant information such as the content of evaluation texts fed back by the user and the specific scoring of the commodity by the user according to the feedback relationship and the occurrence time of the commodity by the user on the graph and by utilizing the attribute of the commodity node. Taking the SSG model (Set-Sequence-Graph:A Multi-View Approach Towards Exploiting Reviews for Recommendation.In CIKM.2020.) as an example, the completion operation of the evaluation scoring matrix is carried out. SSG is the latest work for comprehensively utilizing the attributes of the user and commodity nodes, the feedback relation and the feedback action occurrence time and the content of the user published evaluation text to predict commodity scoring of the user. The method utilizes multi-view learning to evaluate the user from three views: the view angle, the sequence view angle and the view angle are collected to be seen, and further, the representation vectors of users, commodities and feedback are learned, and the scoring of the target 'user-commodity' pair is further predicted. Wherein, the set view angle is to consider different evaluations (or different evaluations under the same commodity) issued by the same user as one set, and then calculate the unified representation of the set as the set representation of the user (or commodity) through the attention network; the sequence visual angle is to arrange the evaluations (or different evaluations under the same commodity) issued by the same user into a sequence according to the time sequence, and then calculate the unified representation of the sequence as the sequence representation of the user (or commodity) through LSTM (long short term memory network); the view angle of the graph utilizes the network relation between the user and the commodity to enable the node and the edge to fully interact with each other through RGAT (drawing meaning network for introducing evaluation), and comprehensively learns the representation vectors evaluated on the user, the commodity node and the edge on the graph as the representation under the view angle of the graph. The three visual angle representations are combined according to a certain proportion through super parameters to predict the commodity scoring of the end user.
4) And performing personalized evaluation generation by using a personalized evaluation text generation model, and simultaneously maintaining synchronous optimization of evaluation scoring prediction tasks. The invention can fully utilize the evaluation scoring prediction and the evaluation text content to generate the connection between the two tasks, so that the two tasks are mutually promoted. Furthermore, the invention can also use a transducer architecture to introduce user interest representation into the transducer for more accurate and more personalized evaluation text content generation.
5) Attack detection task and interaction with the first two tasks (evaluation scoring prediction task and personalized evaluation text generation task), we can take a method similar to GraphRfi(GCN-Based User Representation Learning for Unifying Robust Recommendation and Fraudster Detection.In KDD.2019.) and expand. Specifically, graphRfi models are divided into two tasks, evaluation scoring prediction and water army detection, which are respectively learned by a graph convolution network (Graph Convolutional Networks) and a neural random forest (Neural Random Forests), and are jointly trained. In the evaluation scoring task, if the evaluation scoring of a user is significantly different from the predicted value, the user is highly likely to be a water army user; in a water force detection task, if a user is classified as a water force, the confidence level of the user may be low, and the influence of the user in an evaluation scoring task needs to be reduced. As described above, graphRfi is inadequate in utilizing user feedback information, and does not utilize specific text content to the user's assessment of the merchandise release. According to the invention, evaluation text information is further introduced on the basis of GraphRfi, and a personalized evaluation text generation task is added to perform combined training with other two tasks so as to promote each other. In this component, we consider the first two tasks (the ratings prediction task and the personalized ratings text generation task) as an overall ratings module, analogized to the ratings module in GraphRfi, as shown in fig. 1. The scoring module further interacts with the attack detection task.
Based on the same inventive concept, another embodiment of the present invention provides a robust recommendation system for injection attack resistance in combination with an evaluation text, comprising:
The model training module is used for training a robust recommendation model for resisting injection attack by combining the evaluation text by adopting the method;
And the recommendation module is used for predicting the corresponding evaluation scoring of the user-commodity to be predicted by using the robust recommendation model of the injection attack resistance combined with the evaluation text after the training is finished, and recommending the commodity according to the evaluation scoring.
Based on the same inventive concept, another embodiment of the present invention provides an electronic device (computer, server, smart phone, etc.) comprising a memory storing a computer program configured to be executed by the processor, and a processor, the computer program comprising instructions for performing the steps in the inventive method.
Based on the same inventive concept, another embodiment of the present invention provides a computer readable storage medium (e.g., ROM/RAM, magnetic disk, optical disk) storing a computer program which, when executed by a computer, implements the steps of the inventive method.
The present invention is not limited to the manner described in the above embodiments, such as:
1. The statistics reflecting whether the user is abnormal or not can be limited to the total number of commodities evaluated by the user, the proportion of scoring 1 to 5 points by the user, the entropy value scored by the user, the proportion of all the evaluations of the user being "praised", the interval time of feedback behaviors of the user and the like, and can also comprise any other statistics which are calculated according to node attributes, side structures and side timestamps on the graph and can reflect the water force characteristics of the user;
2. Based on the information on the existing graph, a model for performing evaluation scoring prediction may use other graph neural network or non-graph neural network models, including other graph representation learning models that partially utilize the information on the graph (underutilization of the information on the existing graph).
3. Other text generation models such as LSTM, transformer and expansion thereof can be adopted, and interaction and co-optimization of personalized text generation tasks and evaluation scoring tasks can be carried out in other modes besides multi-task learning and two-way learning.
4. The co-optimization of these three tasks may be performed in a manner that is interactive with three other tasks than illustrated by the present invention.
The above embodiments are only for illustrating the technical solution of the present invention and not for limiting the same, and those skilled in the art may modify or substitute the technical solution of the present invention without departing from the principle and scope of the present invention, and the protection scope of the present invention shall be defined by the claims.
Claims (8)
1. The robust recommendation model training method for resisting injection attack combined with evaluation text is characterized by comprising the following steps of:
training an evaluation scoring prediction model;
training a personalized evaluation text generation model according to the evaluation scoring predicted by the evaluation scoring prediction model and the generated user representation and commodity representation;
Training an attack detection model according to errors of a predicted value and a true value of the evaluation scoring prediction model and the personalized evaluation text generation model, and detecting injection evaluation;
according to the injection evaluation detection result output by the attack detection model, reducing the influence weight of the corresponding evaluation in the evaluation scoring prediction model and the personalized evaluation text generation model;
The training evaluation scoring prediction model is obtained by training a graph neural network according to user attributes, commodity attributes, evaluation contents of commodities given by a user, scoring and feedback time generated by the user to obtain representations of the user, the commodity and feedback edges, and the evaluation scoring prediction model is obtained; the user attribute is a statistic capable of reflecting whether a user is abnormal or not;
The method comprises three tasks: an evaluation scoring prediction task realized by using the evaluation scoring prediction model, a personalized evaluation text generation task realized by using the personalized evaluation text generation model, and an attack detection task realized by using the attack detection model; the three tasks interact, promote each other: a personalized evaluation text generation task assists in evaluating a scoring prediction task; the evaluation scoring prediction task guides the personalized evaluation text generation task; in the attack detection task, if one evaluation is classified as an injection evaluation, the influence of the evaluation in the first two tasks is reduced;
for the personalized evaluation text generation task, training a personalized evaluation text generation model cg (U, I, C, R, T) to C 'so that the generated evaluation C' of the edge to be tested is close to the real evaluation Wherein U represents a user node set, I represents a commodity node set, C represents an evaluation text content set, R represents a scoring set of commodities given by a user, T is a time stamp set, and C' is a generated evaluation of an edge to be tested; adding a predictive score re' generated by an evaluation scoring predictive model to conduct guidance when the personalized evaluation text generation model is trained, and reducing the influence weight of false evaluation content according to the output of the attack detection model when the personalized evaluation text generation model is trained; at the same time, an auxiliary model f for predicting evaluation scoring by evaluation content is introduced, and the user representation and commodity representation generated by using the evaluation scoring prediction model are personalized, the evaluation scoring prediction model and the personalized evaluation text generation model are trained together by using the same set of user representation and commodity representation, so that the evaluation scoring prediction task and the personalized evaluation text generation task are mutually cooperated.
2. The method of claim 1, wherein the three tasks are uniformly end-to-end co-trained using a uniform loss function, the overall loss function being:
L＝Lrating+α·Lcontent_generate+β·Lfraudster
Wherein, L rating is the loss function of the evaluation scoring prediction model, L content_generate is the loss function of the personalized evaluation text generation model, L fraudster is the loss function of the attack detection model, and alpha and beta represent super parameters for controlling three loss ratios.
3. The method of claim 2, wherein the loss function of the scoring predictive model is:
Where p e represents the probability that edge e is true, r' e represents the predictive value of edge e generated by the scoring predictive model, Representing the true score of edge e.
4. The method of claim 2, wherein the step of determining the position of the substrate comprises,
The loss function of the personalized evaluation text generation model is as follows:
Where p e represents the probability that edge e is true, c' e represents the rating of edge e generated by the personalized rating text generation model, Representing a true evaluation of edge e.
5. The method of claim 2, wherein the attack detection model has a loss function of:
Where y e represents a label of whether the edge is real, e| represents the total number of edges, and p e represents the probability that the edge E is real.
6. A robust recommendation method for injection attack resistance in combination with evaluation of text, comprising the steps of:
Predicting the corresponding evaluation scoring of the 'user-commodity' to be predicted by using the robust recommendation model for resisting injection attack combined with the evaluation text after training by adopting the method of any one of claims 1 to 5, and recommending the commodity according to the evaluation scoring.
7. A robust recommendation system for injection attack resistance in combination with evaluation text, comprising:
A model training module for training a robust recommendation model for injection attack resistance of a combined evaluation text by the method of any one of claims 1 to 5;
And the recommendation module is used for predicting the corresponding evaluation scoring of the user-commodity to be predicted by using the robust recommendation model of the injection attack resistance combined with the evaluation text after the training is finished, and recommending the commodity according to the evaluation scoring.
8. An electronic device comprising a memory and a processor, the memory storing a computer program configured to be executed by the processor, the computer program comprising instructions for performing the method of any of claims 1-6.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110018824.1A CN112785331B (en) | 2021-01-07 | 2021-01-07 | Robust recommendation method and system for resisting injection attack by combining evaluation text |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110018824.1A CN112785331B (en) | 2021-01-07 | 2021-01-07 | Robust recommendation method and system for resisting injection attack by combining evaluation text |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN112785331A CN112785331A (en) | 2021-05-11 |
| CN112785331B true CN112785331B (en) | 2024-10-15 |
Family
ID=75756038
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110018824.1A Active CN112785331B (en) | 2021-01-07 | 2021-01-07 | Robust recommendation method and system for resisting injection attack by combining evaluation text |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112785331B (en) |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117473284B (en) * | 2023-11-20 | 2024-08-02 | 灏冉舟网络有限公司 | Three-party transaction platform based on artificial intelligence and method thereof |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108470052A (en) * | 2018-03-12 | 2018-08-31 | 南京邮电大学 | A kind of anti-support attack proposed algorithm based on matrix completion |
| JP2019114053A (en) * | 2017-12-22 | 2019-07-11 | Kddi株式会社 | Recommending apparatus, recommending method and recommending program |
Family Cites Families (15)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20130204833A1 (en) * | 2012-02-02 | 2013-08-08 | Bo PANG | Personalized recommendation of user comments |
| KR101437502B1 (en) * | 2012-06-04 | 2014-09-05 | (주)네오위즈게임즈 | Method, apparatus, and recording medium implementing mobile application marketing |
| KR101573601B1 (en) * | 2014-03-10 | 2015-12-04 | 단국대학교 산학협력단 | Apparatus and method for hybrid filtering content recommendation using user profile and context information based on preference |
| CN105574003B (en) * | 2014-10-10 | 2019-03-01 | 华东师范大学 | A kind of information recommendation method based on comment text and scoring analysis |
| JP6566515B2 (en) * | 2015-07-24 | 2019-08-28 | 大学共同利用機関法人情報・システム研究機構 | Item recommendation system and item recommendation method |
| US10904360B1 (en) * | 2015-12-02 | 2021-01-26 | Zeta Global Corp. | Method and apparatus for real-time personalization |
| JP2018063484A (en) * | 2016-10-11 | 2018-04-19 | 凸版印刷株式会社 | User evaluation prediction system, user evaluation prediction method, and program |
| US10963941B2 (en) * | 2017-09-08 | 2021-03-30 | Nec Corporation | Method and system for combining user, item and review representations for recommender systems |
| CN108304479B (en) * | 2017-12-29 | 2022-05-03 | 浙江工业大学 | Quick density clustering double-layer network recommendation method based on graph structure filtering |
| CN111414549A (en) * | 2019-05-14 | 2020-07-14 | 北京大学 | An Intelligent Universal Evaluation Method and System for Recommendation System Vulnerability |
| CN110210933B (en) * | 2019-05-21 | 2022-02-11 | 清华大学深圳研究生院 | Latent semantic recommendation method based on generation of confrontation network |
| CN111222332B (en) * | 2020-01-06 | 2021-09-21 | 华南理工大学 | Commodity recommendation method combining attention network and user emotion |
| CN111275521B (en) * | 2020-01-16 | 2022-06-14 | 华南理工大学 | Commodity recommendation method based on user comment and satisfaction level embedding |
| CN111639184A (en) * | 2020-06-01 | 2020-09-08 | 复旦大学 | Detection system for tendency inconsistency of scores and comment contents |
| CN111930926B (en) * | 2020-08-05 | 2023-08-29 | 南宁师范大学 | Personalized recommendation algorithm combined with comment text mining |
-
2021
- 2021-01-07 CN CN202110018824.1A patent/CN112785331B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2019114053A (en) * | 2017-12-22 | 2019-07-11 | Kddi株式会社 | Recommending apparatus, recommending method and recommending program |
| CN108470052A (en) * | 2018-03-12 | 2018-08-31 | 南京邮电大学 | A kind of anti-support attack proposed algorithm based on matrix completion |
Also Published As
| Publication number | Publication date |
|---|---|
| CN112785331A (en) | 2021-05-11 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Dogan et al. | Are clean energy and carbon emission allowances caused by bitcoin? A novel time-varying method | |
| US20200202071A1 (en) | Content scoring | |
| Taboga | Under‐/Over‐Valuation of the stock market and cyclically adjusted earnings | |
| Jiang et al. | Social contextual recommendation | |
| Nagengast et al. | The great collapse in value added trade | |
| US20170221111A1 (en) | Method for detecting spam reviews written on websites | |
| CN108665339B (en) | E-commerce product reliability index based on subjective emotion measure and implementation method thereof | |
| KR20180041174A (en) | Risk Assessment Methods and Systems | |
| AU2011205137A1 (en) | Social media variable analytical system | |
| CN101645066B (en) | A method for monitoring novel words on the Internet | |
| Buncic et al. | An estimated New Keynesian policy model for Australia | |
| CN111966915B (en) | Information inspection method, computer equipment and storage medium | |
| CN119250963B (en) | A commercial credit assessment and supervision method based on multimodal co-evolutionary algorithm | |
| CN119719319B (en) | AI intelligent customer service response method and system based on remote digital service | |
| CN108446291A (en) | The real-time methods of marking and points-scoring system of user credit | |
| WO2014105940A1 (en) | Techniques for measuring video profit | |
| Qayyum et al. | FRD-LSTM: a novel technique for fake reviews detection using DCWR with the Bi-LSTM method | |
| CN113434628B (en) | Comment text confidence detection method based on feature level and propagation relation network | |
| Qiu et al. | Crowdeval: A cost-efficient strategy to evaluate crowdsourced worker's reliability | |
| Djogbenou | Model selection in factor-augmented regressions with estimated factors | |
| Aliaj et al. | Nowcasting inflation with Lasso‐regularized vector autoregressions and mixed frequency data | |
| CN112785331B (en) | Robust recommendation method and system for resisting injection attack by combining evaluation text | |
| Souza et al. | Evaluating stream classifiers with delayed labels information | |
| CN118193824A (en) | Public opinion heat prediction method | |
| He | Automatic Quality Assessment of Speech‐Driven Synthesized Gestures |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |