JP2016062411A - Information analysis device, information analysis method, and information analysis program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide an information analysis device, information analysis method, and information analysis program capable of performing effective machine learning.SOLUTION: A server device 20 that is an information analysis device includes: log acquisition means 31 that obtains action logs at the time when a user identified by a user ID accesses a service providing site by using a user terminal; clustering means 32 that classifies action patterns based on action logs obtained by the log acquisition means 31 into at least one or more clusters; and machine learning means 34 that generates an identification function by performing machine learning using a plurality of identities including cluster information obtained by classification by the clustering means 32.SELECTED DRAWING: Figure 2

Description

  The present invention relates to an information analysis apparatus, an information analysis method, and an information analysis program that perform machine learning from a user access log or the like.

An ID value evaluation apparatus that evaluates a user's value in consideration of the relationship between the user and the service in order to determine a remuneration amount for an advertisement publisher placed on a WEB site is known (see, for example, Patent Document 1). ).
The ID value evaluation apparatus of Patent Literature 1 generates an evaluation function by machine learning using a set of user attribute information, service attribute information, general information, and user action history prepared in advance as learning data, and evaluation is performed using this evaluation function. A value evaluation value indicating the value of the target user was output.

Here, the user attribute information is, for example, information such as the user's age and gender, and the service attribute information is advertisement site attribute information such as browsing site type, popularity, advertising product, price, etc. Site attribute information. The general information is information independent of user and service attributes such as date and time, season, special event, and event venue.
The user's action history is data that is given qualitatively depending on whether or not a purchase has been made and whether or not the information has been introduced on a blog.

JP 2012-48360 A

  The machine learning as in Patent Literature 1 can be used to provide services such as advertisements and contents that are highly effective for users when performing advertisement distribution and content distribution to the users. For this reason, there is a need for an information analysis apparatus and an information analysis method that can provide services such as advertisements and contents that are more effective for each user by performing more effective machine learning.

  An object of the present invention is to provide an information analysis apparatus, an information analysis method, and an information analysis program capable of performing effective machine learning.

  The information analysis apparatus of the present invention is based on log acquisition means for acquiring an action log when a user specified by a user ID accesses a service providing site using a terminal, and the action log acquired by the log acquisition means Clustering means for classifying action patterns into at least one cluster and machine learning means for generating a discrimination function by performing machine learning using a plurality of features including cluster information classified by the clustering means It is characterized by.

  According to the present invention, the clustering unit classifies the behavior pattern of the behavior log acquired by the log acquisition unit into at least one cluster. The machine learning means generates an identification function using the cluster information classified by the clustering means as a feature. For this reason, when performing machine learning, it is possible to learn by associating the features of a user who performs a similar behavior pattern, so that effective machine learning can be performed.

It is a block diagram which shows the structure of the information analysis system which concerns on embodiment. It is a block diagram which shows the structure of the server apparatus which concerns on embodiment. It is a figure which shows an example of the access log which concerns on embodiment. It is a flowchart which shows the information analysis process which concerns on embodiment. (A) is a diagram showing the transition of the number of accesses per hour, (B) is a diagram showing binarized behavior patterns indicating the presence or absence of access every hour, (C) is a grouping of similar patterns It is a figure which shows an action group. It is a figure which shows the log data which show the presence or absence of the access for every hour for every user ID. It is a figure which shows the log data normalized. (A) is a figure which shows the clustering number for every user ID, (B) is a figure which shows the action pattern for every clustering number. It is a figure which shows category ID and its content. It is a figure which shows an example of a text advertisement. It is a figure which shows the data of the SVM format input into a machine learning means. It is a flowchart which shows the process which delivers an advertisement by a CTR prediction model.

Hereinafter, embodiments of the present invention will be described with reference to the drawings.
[overall structure]
FIG. 1 is a block diagram showing a schematic configuration of an information analysis system 1 of the present embodiment.
As shown in FIG. 1, the information analysis system 1 includes a user terminal 10 and a server device 20 that functions as the information analysis device of the present invention. The user terminal 10 and the server device 20 are connected to a network (for example, the Internet or the like). Communication via a WAN (Wide Area Network) 5.

[User terminal configuration]
Examples of the user terminal 10 include various devices such as a desktop personal computer, a notebook personal computer, a smartphone, a mobile phone, and a tablet terminal. Further, as shown in FIG. 1, a single user may have a plurality of devices such as a desktop personal computer (PC 10A), a smartphone 10B, and a tablet terminal 10C.

A schematic configuration of a general user terminal 10 will be described with reference to a block diagram showing a configuration of a user terminal 10D shown in FIG.
As shown in FIG. 1, the user terminal 10 includes a display unit 11, an input operation unit 12, a terminal communication unit 13, a terminal storage unit 14, and a terminal control unit 15.

The display unit 11 is configured by a liquid crystal display, for example, and displays a predetermined image under the control of the terminal control unit 15.
The input operation unit 12 outputs an operation signal generated by a user operation to the terminal control unit 15. Examples of the input operation unit 12 include a touch panel provided integrally with the display unit 11, an input device such as a keyboard and a mouse, and the like.
The terminal communication unit 13 communicates with a predetermined device on the network 5.

The terminal storage unit 14 is configured by a data recording device such as a memory or a hard disk. The terminal storage unit 14 stores, for example, a user ID for specifying a user, device data for determining a device of the user terminal 10, and the like.
As the user ID, a number stored in a cookie transmitted from the server device 20 when the user terminal 10 accesses the server device 20 can be used.
As the device data, for example, the device manufacturer of the user terminal 10, the device model name, the carrier name of the communication company, and the like are recorded.
In addition, the terminal storage unit 14 stores various programs for controlling the user terminal.

  The terminal control unit 15 includes an arithmetic circuit such as a CPU (Central Processing Unit) and a storage circuit such as a RAM (Random Access Memory), and controls each unit of the user terminal 10. The terminal control unit 15 expands a program (software) such as a predetermined application from among the programs stored in the terminal storage unit 14 and the like, and performs various processes in cooperation with the program expanded in the RAM. Run. Thereby, the terminal control unit 15 can communicate with the server device 20 via the network 5, and can use various services provided by the server device 20, for example. At this time, the terminal control unit 15 transmits device data for specifying the user terminal 10, a user ID for specifying the user, and the like to the server device 20 at the same time.

[Configuration of server device]
FIG. 2 is a block diagram illustrating a schematic configuration of the server device 20.
The server device 20 is configured by a computer and functions as an information analysis device of the present invention, and also functions as a service providing device such as a WEB server that provides various services to the user terminal 10. The server device 20 includes a server communication unit 21, a server storage unit 22, a server control unit 30, an input device 24, a display device 25, and the like.

The server communication unit 21 is connected to the network 5 and communicates with the user terminal 10.
The input device 24 includes a keyboard, a touch panel, a mouse, and the like. The input device 24 is used when inputting analysis conditions and the like. For example, hyperparameters α and β for clustering, the upper limit number of clustering, and the like can be input with the input device 24.
The display device 25 is composed of a liquid crystal display or the like, and displays analysis results and the like. The server device 20 is not provided with the input device 24 and the display device 25, but an analysis terminal is connected to the server device 20 via a LAN or the like, and an analysis condition is input at the analysis terminal or an analysis result is obtained. May be displayed.

  The server storage unit 22 is a data recording device configured with, for example, a memory, a hard disk, and the like. The server storage unit 22 accumulates and stores an access log as an action log (action history data) of the user terminal 10. That is, the server storage unit 22 functions as an action log storage unit. The server storage unit 22 stores various programs for controlling the server device 20.

[Configuration of server control unit]
The server control unit 30 includes an arithmetic circuit such as a CPU and a storage circuit such as a RAM. The server control unit 30 expands a program (software) stored in the server storage unit 22 and the like in the RAM and cooperates with the program expanded in the RAM. Various processes are executed by operation. Then, the server control unit 30 executes the above-described various processes, and as shown in FIG. 2, the log acquisition unit 31, the clustering unit 32, the representative pattern generation unit 33, the machine learning unit 34, the evaluation value calculation unit 35, It functions as information distribution means 36, cluster characteristic estimation means 37, service provision means 40, and the like.

[Log acquisition means]
The log acquisition unit 31 acquires a user ID, an access time (time information), device data of the user terminal 10, use service data as an access destination service, and the like when accessed from the user terminal 10. In association with each other, it is stored in the server storage unit 22 as an access log as shown in FIG. That is, the access log includes time information indicating the time when the access behavior of the user (user terminal 10) to be analyzed is detected, and classification information (device data and usage service data) that can identify the user's situation. Action log (action history) data included. Therefore, the log acquisition unit 31 functions as an action history acquisition unit that acquires action log data (access log).

The user ID is an anonymized ID that indicates a user to be analyzed. When the same user has a plurality of devices, a user ID may be set for each user, but it is preferable to set a different user ID for each different device of the same user. Thereby, an action log can be accumulated for every device used, and a user's action log can be acquired in detail.
The access time is the time when the user terminal 10 accesses the server device 20 that is a WEB server.

The device data included in the access log is the device type of the user terminal 10 as described above. This device data is data that can be selected from one of multiple values. That is, the device data is a parameter for selecting one of four values such as the personal computer 10A, the smartphone 10B, the tablet terminal 10C, and other terminals. In the above example, the device data shows an example of any of four values, but for example, a value when there is an access from a wearable terminal that is worn and used by the user may be added. Furthermore, when the device is a smartphone, it may be data that further classifies the model, whether it is for work or private use.
Note that the device data is not necessarily essential data, but if the device data is provided, the access environment can be analyzed in detail.

  The usage service data included in the access log is also data that can be selected from any of multiple values. In other words, the usage service data is, for example, a parameter indicating which of the services provided by the service providing unit 40 is accessed, and is a multi-value parameter corresponding to the number of services provided. For example, when 10 types of services such as mail, headline news, daily news, auction, shopping, finance, wisdom bag, baseball, map, etc. are set as services to be classified, the usage service data is the 10 It is set to one of 10 values indicating that one of the types of services has been accessed.

  The classification information included in the access log is not limited to device data and usage service data, but may be location data indicating an accessed place, weather data such as weather, temperature, and humidity at the time of access. In short, the type of classification information may be set according to the purpose of analyzing the behavior pattern.

[Clustering means]
The clustering means 32 classifies each user's action pattern based on the access log stored and accumulated in the server storage unit 22 into at least one or more action groups (clusters) using a Dirichlet process or the like. (Clustering).
Therefore, the clustering unit 32 includes an analysis target setting unit 321, an action history data extraction unit 322, an action pattern creation unit 323, and a grouping unit 324. Details of these means will be described in the clustering process described later.

[Representative pattern generation means]
The representative pattern generation unit 33 generates a representative behavior pattern for each behavior group.
Therefore, in this embodiment, the behavior pattern analysis apparatus is configured by the clustering unit 32 and the representative pattern generation unit 33.

[Machine learning means]
The machine learning unit 34 generates a discriminant function f (x) using a technique such as a support vector machine (SVM). The discriminant function f (x) is expressed as the following formula (1).

  Here, x represents each feature, w represents a weight vector of each feature, b represents a bias term, and d represents the number of dimensions. Specifics of an identification function f (x) used for predicting CTR for a text advertisement distributed / displayed on a page (for example, a news page) accessed on the user terminal 10 An example is shown in equation (2).

The features used in equation (2) appear in both the actual CTR, the cosine similarity of two types of words and category IDs, the AD (advertisement) word weight, and both AD and user attributes. The weights of the category IDs to be received, the weights of the positions for each domain of the page, and the clustering result and the cross feature combining the categories, which are features newly added in the present invention, are used.
Specifically, x _NCTR and w _NCTR are the features and weights of the actual CTR. The actual CTR is a value obtained by dividing the number of times the advertisement has been clicked in the past by the number of times of distribution. The actual CTR can be aggregated with various granularities. For example, the unit of the advertisement, the unit of the advertiser, the gender and age of the user, or the unit obtained by multiplying them.
x _cos and w _cos are the features and weights of cosine similarity of words. The word cosine similarity is based on the similarity between the page accessed by the user terminal 10 and the advertisement, and the cosine similarity between the word vector of the page and the word vector of the text advertisement is obtained.
x _btcos and w _btcos are the features and weights of the cosine similarity of the category. The category cosine similarity refers to the degree of association between the user and the advertisement. The category is assigned to both the user and the advertisement, and the cosine similarity of the category vector is obtained.
_xtermα and _wtermα are the word features and weights of the advertisement (AD).
x _fidi and w _fidi are the features and weights of categories that appear (co-occur) in both advertisements (AD) and users.
w _{rk, d} is the weight of the position for each domain of the page.

x _new and w _new are features and weights obtained by multiplying the cluster information and the category information, which are features of the present invention. The cluster information is a number representing which of the clusters (behavior patterns) that the action pattern of the access operation by the user terminal 10 accessing the page is classified and grouped by the clustering means 32. For example, when the behavior pattern clusters are classified into eight types, the behavior pattern of the user terminal 10 is set to any number of cluster information “1” to “8”.
The category information is a code that represents the category of each distributed advertisement.
In x _new , a cross feature obtained by passing the cluster information and category information is set.
The machine learning means 34 calculates the weight of each feature by performing machine learning.

[Evaluation value calculation means]
The evaluation value calculation means 35 obtains a predicted CTR that is a probability that each text advertisement is clicked as an evaluation value. Specifically, the evaluation value calculation means 35 calculates a predicted CTR, which is an evaluation value, using a CTR prediction model of the following equation (3) constructed using a logistic regression model.

  In the equation (3), f is an identification function of the equation (2) obtained by the machine learning means 34, and A and B are constants.

[Information distribution means]
The information distribution unit 36 distributes information (text advertisement) having a high evaluation value (predicted CTR) calculated by the evaluation value calculation unit 35 to a page accessed by the user.
The information distribution means 36 distributes an advertisement with a high predicted CTR in order to distribute and display an advertisement with a high expected profit. In other words, in the click charge type advertisement, the advertiser pays the price only when the user clicks the advertisement. Therefore, in order to maximize the profits that the online advertising service provider can obtain, it is necessary to distribute advertisements with high expected profits (eCPM) to user terminals.
In order to obtain an accurate expected profit (eCPM), it is necessary to predict CTR. Expected revenue is obtained from the bid amount (Bid) set by the advertiser and the click-through rate (CTR) (eCPM = CTR × bid).
Here, when an advertisement is displayed on a user terminal operated by a certain user, the probability of clicking the advertisement with a certain probability is unknown, so it is necessary to predict it. For this reason, the CTR prediction model of Expression (3) is constructed by machine learning using Expression (2), and the information distribution means 36 distributes the advertisement with high predicted CTR with priority.

[Cluster characteristic estimation means]
The cluster characteristic estimation unit 37 estimates the characteristics of each cluster grouped by the clustering unit 32. That is, since the clustering means 32 performs clustering using a Dirichlet process or the like, it is unclear what characteristics each grouped cluster has. Therefore, as will be described later, the cluster characteristic estimation unit 37 uses the cross feature obtained by passing the cluster information and the category information in the machine learning unit 34, and checks each cross feature having a high weight. Estimate cluster characteristics.

[Service provision means]
The service providing unit 40 provides the user terminal 10 with a service that can be used on the WEB in response to a request from the user terminal 10. The services provided by the service providing means 40 include, for example, information transmission services such as news, online commercial transaction services (for example, stock transactions and online shopping), mail services, and the like.

[Information analysis method]
Next, a processing method in the server apparatus (information analysis apparatus) 20 as described above will be described with reference to the flowchart of FIG. Steps S1 to S4 in FIG. 4 are clustering processing for classifying the behavior groups by clustering the behavior patterns of each user based on the accumulated access log, and step S5 is representative pattern generation processing for generating the representative patterns. It is. Further, step S6 is a process for creating a cross feature, and step S7 is a process for machine learning using a pairwise method or the like.

Details of each process will be described below.
When the server device 20 is accessed from the user terminal 10 to a predetermined service providing site, the server device 20 detects and accumulates the access log (behavior log).
Services provided by the service providing site include information transmission services such as news, online commerce services (for example, stock trading and online shopping), mail services, and the like. These service providing sites can be constructed by the server device 20 or by other server devices not shown.
Furthermore, the server device 20 can accumulate access logs for each service providing site, or can collect access logs for these various services collectively.

[Clustering processing]
Clustering processing by the clustering means 32 of the server device 20 will be described based on the drawings.
As an analysis preparation, the analysis target setting unit 321 of the clustering means 32 sets an analysis target condition for designating what kind of user behavior is analyzed when analyzing the behavior pattern (step S1). Specifically, collection period information for collecting behavior log data to be analyzed, analysis period information indicating a period for analyzing a behavior pattern, and unit period information for analyzing a behavior pattern are set. Further, the hyper parameters α and β and the upper limit number of clustering are appropriately input. For example, both the hyper parameters α and β are set to “1”, and the upper limit number of clustering is set to “10”. This setting is set when an analyst inputs to the server device 20 using the input device 24.

The collection period information sets the analysis target period of the access log stored in the server storage unit 22. For example, period information is set at the beginning and end of the collection period, such as from May 11 to August 18.
The analysis period information sets what period of behavior pattern is to be analyzed. For example, when analyzing a daily behavior pattern, the analysis period information is set to one day (for example, 00:00 to 23:59:59). Further, when analyzing the behavior pattern for each week, the analysis period information is set to one week (for example, Sunday to Saturday). Furthermore, when analyzing a behavior pattern every 6 hours, the analysis period information is 6 hours (for example, from 00:00:00 to 05:59:59, 06:00:00 to 11:59:59, 12 : 00: 00 to 17:59:59, 18:00:00 to 23:59:59).
The unit period information is a time interval shorter than the analysis period information, and sets a unit time for determining the presence or absence of an action (access). For example, when the analysis period information is one day, the unit period information is an hour unit (for example, 00:00:00 to 00:59:59, 01:00 to 01:59:59,..., 23 : 00: 00 to 23:59:59). The unit period information is not limited to one hour unit, may be a unit of 30 minutes, may be a unit of two hours, and may be set according to the purpose of analyzing the behavior pattern.

The classification information is used to set classification information to be analyzed from the accumulated access log, that is, under what circumstances the behavior pattern of the accessed user is analyzed. For example, device data or service data can be specified as the classification information. It is also possible to specify a plurality of classification information such as device data and service data. Therefore, only the access log accessed to the news site can be the target of analysis (clustering).
Further, one or more devices in the device data can be designated as the classification information. For example, only the personal computer 10A may be specified, only the smartphone 10B may be specified, or two devices of the smartphone 10B and the tablet terminal 10C may be specified. Similarly, one or more services of service data may be specified. Therefore, the classification information is set by the analyst according to the analysis purpose.

Next, based on the collection period information, analysis period information, and unit period information set by the analysis target setting unit 321, the action history data extraction unit 322 selects action history data (action log) corresponding to the analysis target condition from the access log. Data) is extracted (step S2). For example, if the collection period information is from May 11 to August 18, the analysis period information is every day, and the unit period information is 1 hour, the action history data extracting unit 322 may access time to the collection period. The behavior log data is acquired, the acquired data is divided every day, the behavior log data for each day is extracted, and the behavior log data for each day is totaled every hour.
FIG. 5A shows the action log data for the extracted user for one day. At this time, for example, the news site is specified as the service data to be used for the analysis target condition classification information, and the device data is not specified. The number of accesses is counted every hour. For this reason, the graph of FIG. 5A is a graph showing the transition of the number of accesses for each hour with the horizontal axis representing the time of the day and the vertical axis representing the number of accesses.

  Next, the behavior pattern creation unit 323 detects whether or not the server device 20 has been accessed every unit period (one hour unit) from the extracted behavior log data. If not, binarized data that returns “0” is created as an action pattern (step S3). That is, the behavior pattern creation means 323 determines the transition of the number of accesses per day for each user shown in FIG. 5 (A) for each category information (in the example of FIG. The behavior pattern graph shown in FIG. 5B is created by dropping into the distribution. In addition, if the action pattern of FIG. 5 (B) is represented by numerical data, it will become as shown in FIG. That is, the log data represents the user ID and the presence / absence of access in each time zone from 0:00 to 23:00 as “1” and “0”.

The grouping means 324 executes a script for grouping (clustering) similar patterns among the user's action patterns created by the action pattern creating means 323 (step S4).
Note that the action log data of each user input to the grouping unit 324 is not limited to data that can be represented by binary values of “1” and “0”. For example, as shown in FIG. It may be expressed by a numerical value between. The action log data shown in FIG. 7 is obtained by extracting data having about 1 million user IDs. That is, the number of accesses for each time zone from 0 o'clock to 23:00 of about 1 million user terminals (user IDs) whose access logs are stored for 5 days on weekdays is counted, and the number of accesses for each user ID in each time zone Is normalized by dividing by the total number of accesses of all times of the user ID. Such action log data may be input to the grouping unit 324 to be grouped.
The grouping means 324 performs clustering using these input data. The clustering can be performed by, for example, a non-parametric Bayes method using a composite Dirichlet process model having multi-value parameters.
The grouping means 324 calculates an optimum number of action groups, and outputs a plurality of action groups obtained by clustering action patterns as shown in FIG. FIG. 5C is an example in which four action groups are output.
Further, as shown in FIG. 8A, the grouping unit 324 outputs a label (clustering number, that is, cluster information) indicating which group each user ID corresponds to.

Next, the representative pattern generation unit 33 generates a representative pattern in each action group generated in step S4 (step S5).
In step S5, the representative pattern generation unit 33 calculates, for example, an average value of a plurality of behavior patterns included in each behavior group. FIG. 8B shows an example of a representative pattern.
In FIG. 8B, Type 1 is a type in which the number of accesses is low and constant from morning until night, Type 2 is a type in which the number of accesses increases most around 12:00 am, and Type 3 is a midnight time. This type increases the number of accesses.
Thus, if the representative pattern generation means 33 generates a representative pattern and outputs it to the display device 25 or the like, the analyst can easily grasp the characteristics of each pattern.

The label number (cluster information) that is the output result of the clustering means 32 is used as a feature by the machine learning means 34. In addition, since the machine learning means 34 of this embodiment performs a machine learning by a pairwise method, if the cluster information corresponding to a user ID is simply added, it will be erase | eliminated when the difference is taken by the pairwise method.
Therefore, the machine learning means 34 creates a cross feature that crosses the label number and the category information of the advertisement (step S6).

FIG. 9 is a diagram illustrating an example of advertisement category information. The category ID indicating the category information of the advertisement shown in FIG. 9 and the category contents are associated with each other and stored in advance in the server storage unit 22.
The machine learning unit 34 assigns a new number as a cross feature between the label number 1 to 10 and the category ID when, for example, 10 clusters are generated by the clustering process in step S4.
For example, the cross feature between the category ID “1000738796 (fashion accessory)” and the label number 1 is “10501”, the cross feature between “1000738796 (fashion accessory)” and the label number 2 is “10502”, The cross feature between “1000738796 (fashion accessory)” and label number 10 is “10510”.
In addition, the cross feature between the category ID “1000552812 (various human resource services)” and the label number 1 is “10621”, and the cross feature between “1000552812 (various human resource services)” and the label number 10 is “10630”.
In this way, the machine learning means 34 creates a cross feature by multiplying each category ID and the label number, and assigns a new number.

Next, the machine learning means 34 performs machine learning using the action log data for the set learning period (step S7).
For example, in order to predict a click on a text advertisement as shown in FIG. 10, the machine learning means 34 displays action log data on the presence or absence of a click on each text advertisement in the learning period as shown in FIG. Then, the above equation (2) is input to perform machine learning.
In FIG. 11, one line corresponds to one text advertisement. The first value in each row indicates whether or not there is a click, “+1” indicates that there is a click, and “−1” indicates that there is no click. Therefore, in the example of FIG. 11, since the third line is “+1”, the text advertisement “<<cheap> Kyushu trip” in the third line of FIG. 11 is clicked, and no other text advertisement is clicked. Indicates.
Furthermore, as shown in FIG. 11, action log data for each text advertisement is arranged in each row in the form of feature key: value.
The feature key is a number indicating the type of each feature. For example, in FIG. 11, “1” is “actual CTR”, “2” is “word cosine similarity”, and “3” is “category”. It indicates a preset feature such as “cosine similarity”.
The value of the feature is a value indicating the weight of each feature and is obtained by performing machine learning, thereby generating an identification function (step S8).

  Since the weight of each feature is calculated by performing the above machine learning, the cluster characteristic estimation unit 37 estimates the characteristic of each grouped cluster by examining the cross feature having a high weight (step S9). ). For example, it is assumed that the cross feature between the group having the label number 2, that is, the cluster number 2 and each category ID has the largest weight, for example, the cross feature “10502”. This cross feature “10502” is a cross feature between “1000738796 (fashion accessory)” and label number 2. For this reason, the cluster characteristic estimation means 37 can estimate that the group of cluster number 2 is a group having a characteristic that is interested in fashion accessories. The cluster characteristic estimation means 37 can analyze that the relation between the representative pattern of cluster number 2 and the characteristic of cluster number 2 (fashion accessory which is the category ID field) is high. It can be estimated that the user is interested in the same field, and the information distribution means 36 can effectively distribute information to the user. For example, if the cluster number 2 is an action pattern in which the number of accesses to the news site increases on weekday nights, a fashion / accessory category advertisement is distributed to users who browse the news site on weekday nights. Can be presumed to be effective, and advertising can be distributed effectively.

Next, a method for delivering a selected advertisement using the CTR prediction model using the discriminant function of Expression (2) obtained by the machine learning will be described.
As shown in FIG. 12, when a user accesses a service providing site such as a news site using the user terminal 10, the evaluation value calculation means 35 of the server device 20 includes information on the accessed user and information on the page to be accessed. Are used to search for and extract advertisement candidates to be distributed (step S11).
Then, the evaluation value calculation means 35 predicts the CTR of each advertisement using the CTR prediction model of the formula (3) for the candidate advertisement (step S12).
Next, the information distribution means 36 determines an advertisement to be distributed and its order based on the CTR score order, and distributes the advertisement to the page accessed by the user terminal 10 (step S13).

An action log indicating whether or not the user has clicked on an advertisement distributed using the user terminal 10 is stored in the server storage unit 22 of the server device 20. The machine learning unit 34 performs machine learning of the discriminant function by inputting the action log newly accumulated in the server storage unit 22 in a SVM format at regular intervals (for example, every week), and the accuracy of the discriminant function. Can be improved.
The clustering means 32 may also perform the clustering process at regular intervals using the action log newly accumulated in the server storage unit 22.

[Function and effect]
Since the server device 20 of the present embodiment adds a number (cluster information) obtained by clustering and classifying each user's behavior pattern as the feature of the prediction model, the accuracy of the prediction model can be improved. Actually, when offline evaluation was performed using past logs, the evaluation index of AUC (Area Under the Curve) increased, and it was confirmed that the cluster information was an effective feature. Therefore, when the user actually accesses various service pages using the user terminal 10, an effective advertisement can be distributed, the CTR of the advertisement can be improved, and the expected profit can be improved.

  In addition, when adding cluster information as a feature, it was added as a cross feature combined with a category ID instead of adding the cluster information alone, so the pair-wise method with higher learning accuracy than the point-wise method in machine learning Can be used.

  Furthermore, by performing machine learning using the cross feature, it is possible to determine the weight of each cross feature and determine the characteristics (meaning) of each cluster information group. Conventionally, when an unspecified number of users are clustered and grouped based on their action logs, it has been difficult to determine what attributes the group is grouped with, as in this embodiment By performing machine learning using the cross feature that combines the category ID with the cluster information, it is possible to easily make the meaning to the group of the clustering result, and the analyst can evaluate the cluster from various viewpoints. it can.

The grouping unit 324 of the clustering unit 32 clusters the behavior patterns by the nonparametric Bayes method when clustering the behavior patterns. For this reason, the cluster number of the action group can be automatically set to the optimum value only by setting the upper limit value of the cluster number. For this reason, each user's action pattern can be classified accurately.
Therefore, for example, the number of groups is set in advance to 3 or the like, and the user's behavior pattern is matched to the unique behavior pattern of each user as compared with the case where it is determined whether the user belongs to any of these groups. Optimal clustering can be performed.

Furthermore, in this embodiment, the representative pattern generation unit 33 generates the representative pattern of the action group, so that the analyst can easily analyze the action pattern of each user. In particular, since the cluster characteristic estimation unit 37 can estimate the characteristics of each group based on a cross feature with a high weight, the analyst can easily analyze the user's behavior pattern in association with the field in which the user is interested. For example, a group of users who access a news site on weekday nights can easily analyze such as high interest in fashion.
For this reason, since the behavior pattern of each user can be accurately determined, services and information suitable for each user can be provided to the user, and user convenience can be improved.

[Modification]
In addition, this invention is not limited to embodiment mentioned above, In the range which can achieve the objective of this invention, the deformation | transformation shown below is also included.

[Modification 1]
In the above embodiment, the cross feature combining the cluster information and the category ID is used. However, the feature combined with the cluster information is not limited to the category ID. In other words, any feature other than the feature (for example, age, sex, etc.) having the same value with the same user ID may be used. In other words, what is necessary is just the feature which is not deleted when a difference is taken by the pairwise method.
Furthermore, when performing machine learning by a method other than the pairwise method (for example, the pointwise method), the cluster information may be used as a single feature. In this case, the discriminant function may be, for example, the following equation (4).

[Modification 2]
As the features, position data indicating the current location at the time of access, weather data indicating the weather at the time of access, temperature, humidity, or the like may be used.
The position data is data indicating the current position of the user terminal 10 when the user terminal 10 accesses the server device 20, and may be stored in the access log. For example, when the position data is set to a parameter indicating the prefecture at the current position, the position data is a multi-value parameter corresponding to the number of prefectures.
In addition, when the user terminal 10 includes position measurement means such as a GPS sensor, the position information (latitude and longitude) measured by the position measurement means may be recorded as the position data. Further, when the user terminal 10 does not include a position measuring unit, the position information (latitude and longitude or address data) of the base station with which the user terminal 10 communicates may be recorded as the position data.
If such location data is used as classification information, the situation at the time of access by the user can be classified into classification information according to the access location such as access at home, access at the office, or access during commuting, for example. Based on the behavior pattern can be analyzed. Therefore, each user's action pattern can be classified more accurately.

Furthermore, the weather data may be recorded using data from environmental sensors provided in the base station accessed by the user terminal 10.
If such weather data is used as classification information, it is possible to classify situations at the time of user access, for example, an action pattern when the weather is fine and the humidity is 60% or less. Therefore, it is possible to analyze with high accuracy the behavior pattern of each user and how the CTR is affected by the weather.

[Modification 3]
The evaluation value calculation means 35 calculates the predicted CTR, but the conversion rate (CVR) is the probability of a specific user behavior (for example, product purchase behavior or data request behavior) targeted by the advertiser. ) May be predicted, or another evaluation index may be predicted.

[Modification 4]
Information distributed by the information distribution means 36 is not limited to text advertisements, and may be various contents such as images, moving images, and music.
The representative pattern generation means 33 of the embodiment generates a representative pattern by averaging the average value of each behavior pattern included in the behavior group, that is, the value of “0, 1” of each unit period for each unit period. However, it may be generated by other methods. For example, the representative pattern may be generated with the most frequent value of “0, 1” for each unit period. Furthermore, the similarity of each action pattern within the same group may be calculated, and re-clustering may be performed to collect similar ones, and a representative pattern may be generated based on the midpoint of the similarity of the plurality of action patterns that have been combined. .

  In addition, the specific structure and procedure for carrying out the present invention can be appropriately changed to other structures and the like within a range in which the object of the present invention can be achieved. For example, a computer such as a server having a server storage unit 22 in which the server device 20 is configured as a distributed system and accumulating access logs, a log acquisition unit 31, a clustering unit 32, a representative pattern generation unit 33, a machine learning unit 34, You may provide separately the computer provided with the evaluation value calculation means 35, the information delivery means 36, the service provision means 40 grade | etc.,.

  DESCRIPTION OF SYMBOLS 1 ... Information analysis system analysis system, 10 ... User terminal, 20 ... Server apparatus, 21 ... Server communication part, 22 ... Server memory | storage part, 30 ... Server control part, 24 ... Input device, 25 ... Display apparatus, 31 ... Log acquisition Means 32: Clustering means 321: Analysis target setting unit 322 Action history data extraction means 323 Action pattern creation means 324 Grouping means 33 ... Representative pattern generation means 34 ... Machine learning means 35 ... Evaluation value calculation means, 36... Information distribution means, 37... Cluster characteristic estimation means, 40.

Claims (10)

Log acquisition means for acquiring an action log when a user specified by a user ID accesses a service providing site using a terminal;
Clustering means for classifying the behavior pattern based on the behavior log acquired by the log acquisition means into at least one cluster;
Machine learning means for generating a discriminant function by performing machine learning using a plurality of features including cluster information classified by the clustering means. In the information analysis device according to claim 1,
The information includes a feature that combines an attribute of service content provided on the service providing site and the cluster information. In the information analysis device according to claim 1 or 2,
The information analyzing apparatus, wherein the machine learning means calculates a weight of each feature. In the information analysis device according to any one of claims 1 to 3,
An information analysis apparatus comprising: cluster characteristic estimation means for estimating a characteristic of the cluster based on a feature weight obtained by combining an attribute of service content provided at the service providing site and the cluster information. In the information analysis device according to any one of claims 1 to 4,
The log acquisition means acquires an action log including time information when the user accesses a service providing site using a terminal,
The clustering means extracts the behavior log for each predetermined analysis period, divides the analysis period into specific time units, and creates the behavior pattern based on a value indicating a user access state in each time unit. And classifying the behavior pattern into at least one cluster according to the similarity of the behavior pattern. In the information analysis device according to any one of claims 1 to 5,
Evaluation value calculating means for calculating an evaluation value corresponding to the user ID using the identification function;
An information analysis apparatus comprising: information distribution means for distributing distribution information selected based on the evaluation value to the service providing site. The information analysis device according to claim 6,
The distribution information distributed to the service providing site is an advertisement or content,
The evaluation value calculating means calculates a predicted CTR that predicts a probability that the delivery information is clicked,
The information distribution unit distributes distribution information having a high predicted CTR to the service providing site. In the information analysis device according to claim 7,
The feature used in the machine learning means is
A track record CTR based on a past track record of the distribution information;
Similarity between the attribute of the service providing site and the attribute of the distribution information;
The similarity between the user attribute identified by the user ID and the attribute of the distribution information;
Attributes of the delivery information;
An information analysis apparatus comprising: a position indicating a display position of the distribution information. An information analysis method for performing machine learning processing by a computer,
The computer
Acquire an action log when the user specified by the user ID accesses the service providing site using the terminal,
Classifying action patterns based on the acquired action logs into at least one cluster;
A discriminant function is generated by performing machine learning using a plurality of features including the classified cluster information. An information analysis program read and executed by a computer,
An information analysis program that causes the computer to function as the information analysis apparatus according to any one of claims 1 to 8.

Images