JP2018073349A - Information analysis device, information analysis method, and information analysis program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To further adequately determine a probability for a purchase action.SOLUTION: An information analysis device comprises: an acquisition unit for acquiring a purchase history of a user to which at least one incentive out of a plurality of incentives having different influence degrees on the user from one another is given, and identity information on each user to which the incentive is given; a model generation unit for generating, on the basis of the purchase history of the user and the identity information on each user acquired by the acquisition unit, a prediction model made by modeling a purchase trend for the user for each incentive; and a probability prediction unit for predicting, on the basis of the prediction model generated by the model generation unit, a probability for optimizing an event related to a purchase action occurring by giving the user at least one incentive out of the plurality of incentives.SELECTED DRAWING: Figure 1

Description

  The present invention relates to an information analysis device, an information analysis method, and an information analysis program.

  2. Description of the Related Art Conventionally, a technique for giving incentives to a more effective user from a use history of incentives such as coupons and points used by the user in the past is known (see, for example, Patent Documents 1 and 2).

JP 2011-243041 A Japanese Patent Laying-Open No. 2006-4557

  However, in the prior art, since the probability of taking an action such as purchase is not obtained under appropriate conditions on the premise that there is an incentive, materials for measuring the effect of the incentive are insufficient.

  The present invention has been made in view of such circumstances, and an object of the present invention is to more appropriately obtain a probability relating to purchase behavior.

  According to one aspect of the present invention, a purchase history of a user to which any one or more incentives are given among a plurality of incentives having different degrees of influence on users, and feature information for each user to which the incentives are given, And a model generation that generates a prediction model that models a purchase tendency by the user for each incentive based on the purchase history of the user and the feature information for each user acquired by the acquisition unit And a probability for optimizing an event related to purchase behavior generated by giving at least one incentive out of the plurality of incentives based on a prediction model generated by the model generation unit An information analysis apparatus including a probability prediction unit.

  According to one aspect of the present invention, it is possible to determine the probability regarding purchase behavior more appropriately.

1 is a diagram illustrating an example of an information analysis system 1 including an information analysis apparatus 100 according to an embodiment. It is a figure showing an example of composition of information analysis device 100 in an embodiment. 5 is a diagram illustrating an example of purchase history information 132. FIG. It is a figure showing other examples of purchase history information 132. It is a figure which shows an example of the feature information. It is the figure which illustrated the feature of user action history more concretely. It is a figure which shows an example of the prediction model information 136. FIG. It is a figure which shows an example of the account information 138. 4 is a flowchart illustrating an example of a series of processes performed by a control unit 110. It is a figure which shows an example of the score derived | led-out for every point. It is a figure which shows the information regarding the purchase history of all the users who used the sales site in a certain period. It is a figure which shows the determination method of the parameter A typically. It is a figure which shows the determination method of the parameter B typically. It is a figure which shows a mode that a future purchase probability is derived | led-out from a score. It is a figure for demonstrating linear programming relaxation. It is a figure which shows an example of the optimal point allocated for every user. It is a figure which shows an example of the hardware constitutions of the information analysis apparatus 100 of embodiment.

  Hereinafter, embodiments of an information analysis apparatus, an information analysis method, and an information analysis program according to the present invention will be described with reference to the drawings.

[Overview]
The information analysis apparatus according to the embodiment is realized by one or more processors. The information analysis device is based on a purchase history of a user to which any one or more incentives are given among a plurality of incentives having different degrees of influence on the user, and feature information for each user to which the incentives are given. Then, a prediction model that models the purchase tendency by the user for each incentive is generated. Then, the information analysis apparatus predicts a probability for optimizing an event related to purchase behavior caused by giving at least one incentive among a plurality of incentives based on the prediction model.

  An incentive is a point that can be used for purchasing goods or services (hereinafter referred to as items) in shopping, for example. Incentives may be coupons, tickets, discount coupons, etc. that have the same effect as points, and the right to preferentially purchase a specific item (a purchase right limited to the first person) Also good. In the present embodiment, as an example, the incentive is described as a point that can be used for purchasing an item.

  Further, the information analysis apparatus realizes effective incentive provision within a limited budget based on the predicted probability.

  The function for predicting the above-described probability and the function for giving an incentive do not need to be realized by the same apparatus, and an information analysis apparatus having only each function may be configured.

<First Embodiment>
[overall structure]
FIG. 1 is a diagram illustrating an example of an information analysis system 1 including an information analysis apparatus 100 according to an embodiment. The information analysis system 1 in the embodiment includes one or more terminal devices 10, one or more server devices 20, and an information analysis device 100. These devices are connected to each other via a network NW. The network NW includes, for example, a wireless base station, a Wi-Fi access point, a communication line, a provider, the Internet, and the like. Note that it is not necessary for all combinations of the devices shown in FIG. 1 to be able to communicate with each other, and the network NW may partially include a local network.

  The terminal device 10 is a terminal device used by a user. The terminal device 10 is, for example, a mobile phone such as a smartphone, a tablet terminal, a PDA, or a personal computer. For example, the terminal device 10 may receive a user's operation and access a website provided by the server device 20 via a predetermined web browser, or with the server device 20 via an application installed in advance. You may communicate and acquire the content (for example, an image, a moving image, a text, etc.) displayed or reproduced on an application. This application may be, for example, an application that can purchase an item such as a shopping application, or a content distribution application that charges within the application and reproduces a moving image or the like.

  The server device 20 may be a web server device that provides websites such as shopping sites, auction sites, and flea market sites (hereinafter collectively referred to as “sales sites”) on the Internet, for example. In this sales site, for example, points can be used at the time of item purchase, and part or all of the purchase amount of the item can be paid by points.

  In addition, the server device 20 may be an application server device that communicates with the terminal device 10 in which the application is started to exchange various types of information. The screen generated by the application program of the terminal device 10 based on information provided by the application server device provides the same service as the sales site. Hereinafter, in order to simplify the description, the server device 20 will be described as a web server device that provides a sales site.

  The information analysis device 100 communicates with one or both of the terminal device 10 and the server device 20 via the network NW, and acquires various types of information regarding the user. Based on the acquired information, the information analysis apparatus 100 predicts the probability of purchase behavior that the user will take when giving points from the purchase tendency of the user who has been given points in the past. And the information analysis apparatus 100 analyzes what should be done in order to give a point more effectively to the user who is going to give a point from now on based on the predicted probability.

[Configuration of information analysis device]
Hereinafter, the configuration of the information analysis apparatus 100 will be described. FIG. 2 is a diagram illustrating an example of the configuration of the information analysis apparatus 100 according to the embodiment. The information analysis apparatus 100 includes a communication unit 102, a control unit 110, and a storage unit 130, for example.

  The communication unit 102 includes a communication interface such as a NIC, for example. The communication unit 102 communicates with other devices such as the terminal device 10 and the server device 20 via the network NW. For example, the communication unit 102 receives terminal-side management information from the terminal device 10. The terminal side management information is, for example, a cookie (HTTP cookie) managed for each web browser. This cookie includes, for example, a purchase history at a sales site and a browsing history of a web page. The communication unit 102 may receive from the server device 20 server-side management information that corresponds to the above-described terminal-side management information and includes purchase history at a sales site, web page browsing history, and the like. The terminal side management information and the server side management information received by the communication unit 102 are stored in the storage unit 130 as purchase history information 132 to be described later.

  The control unit 110 includes, for example, an acquisition unit 112, a model generation unit 114, a purchase probability prediction unit 116, and a grant point determination unit 118. Some or all of these components are realized by a processor such as a CPU (Central Processing Unit) executing a program stored in the storage unit 130, for example. In addition, some or all of the components of the control unit 110 may be realized by hardware such as a large scale integration (LSI), an application specific integrated circuit (ASIC), or a field-programmable gate array (FPGA). It may be realized by cooperation of software and hardware.

  The storage unit 130 is realized by, for example, an HDD (Hard Disc Drive), a flash memory, an EEPROM (Electrically Erasable Programmable Read Only Memory), a ROM (Read Only Memory), or a RAM (Random Access Memory). In addition to the program read by the processor, the storage unit 130 stores information such as purchase history information 132, feature information 134, prediction model information 136, and account information 138.

  FIG. 3 is a diagram illustrating an example of the purchase history information 132. In the purchase history information 132 illustrated in the figure, when each of a plurality of types of incentives (points) is given to a user with respect to a user ID that can identify each user, the user is given a predetermined period (for example, a point validity period). ), Whether you purchased an item that can be purchased using incentives (points) (that is, an item sold on the sales site) ("Purchase" in the figure) or did not purchase ("Non-purchase" in the figure) )) Is associated with the information.

  The user ID is issued by the server device 20 in order to identify each of a plurality of users who use the sales site, for example. The multiple types of incentives have different degrees of influence on the user (for example, the degree to improve the willingness to purchase). For example, 500 points, 300 points, 100 points, 3 points, 0 points, etc. Includes incentives with different effectiveness (usable face value). 0 points means that no points are given. Therefore, the presence / absence of purchase (purchase or non-purchase) at 0 points indicates whether a user who has not been awarded points has purchased a purchaseable item using the points. The “purchase” at each point may or may not be a requirement that the point is used at the time of purchase.

  In the example of the above-mentioned figure, by giving all types of incentives of 500 points, 300 points, 100 points, 3 points, and 0 points to each user ID, the presence / absence of purchase (purchase / non-purchase) in each incentive However, the present invention is not limited to this, and information indicating whether or not a specific incentive has been purchased may be acquired by giving only some incentives. For example, a user ID of a certain population (for example, a set of about 1 million users) is given points different from each other for each predetermined number (for example, 200,000 people), and information indicating the presence or absence of purchase is acquired. It's okay. That is, it is only necessary to determine whether any one of 500 points, 300 points, 100 points, 3 points, and 0 points has been purchased.

  In addition, with respect to an arbitrary user ID (including a user ID that was not used as learning data such as machine learning to be described later) for which information indicating user attributes such as age and gender is obtained, a purchase tendency is modeled for each point. A predictive model can be generated. For example, after generating prediction models at each point using information indicating whether or not 500 points, 300 points, 100 points, 3 points, and 0 points have been purchased for each 200,000 cases, incentives are given after the next time The purchase probability of all points of each user ID is calculated by inputting only the attribute information of each user ID to the prediction model of each point for all user IDs to be processed. Therefore, in the present embodiment, for example, after constructing a prediction model using 1 million learning data in total of 200,000 cases, including user IDs not used for learning, for example, 10 million user IDs are included. The purchase probability can be calculated. Details of these processes will be described later using a flowchart and the like.

  Further, at the stage of acquiring the purchase history information 132, each point is given so that the valid periods do not overlap. For example, after giving 500 points, in order to determine whether or not the item has been purchased by giving 500 points, another type of points is given after waiting for an expiration period of 500 points, for example. In addition, as described above, for example, in the case of granting only a specific point without giving all points to one user, the points are given without considering the expiration period of the points. Good.

  Further, the purchase history information 132 may include information indicating a purchase amount when an item that can be purchased using points is purchased. FIG. 4 is a diagram illustrating another example of the purchase history information 132. In the purchase history information 132 illustrated in the figure, when a user purchases an item when each of a plurality of types of points is given to each user ID, information indicating the purchase price of the item is associated. If the user has not purchased the item, a zero yen is associated. As a result of using points, if an item is purchased for zero yen, a flag indicating that may be given. Similar to FIG. 3 described above, also in FIG. 4, all types of incentives of 500 points, 300 points, 100 points, 3 points, and 0 points are assigned to each user ID. Information indicating whether or not there is a purchase (purchase / non-purchase) need not be acquired, and only information indicating whether or not a specific incentive is purchased may be acquired by giving only some incentives.

  FIG. 5 is a diagram illustrating an example of the feature information 134. As shown in the example of the figure, in the feature information 134, a feature indicating a user attribute, a feature indicating a user behavior history, and a feature related to a point are associated with a user (user ID) to which a point is given. Information.

  The feature indicating the user attribute includes various features such as sex, age (age), registration date of user ID, charge amount of electronic money, and the number of retained points. The registration date of the user ID is the date on which the user ID is issued and registered at the sales site. The charge amount of electronic money is information indicating a charge balance of electronic money that can be used on a sales site or a website affiliated with a sales site. Features indicating the user behavior history include features related to the use of shopping at the sales site, features related to the use of auctions at the sales site, features related to browsing of web pages constituting the sales site, and the like. Features related to points are features related to points that have been given to users in the past, and include, for example, features such as the number of points used, the amount of points used, the number of points earned, the amount of points earned, and the balance of points. .

  FIG. 6 is a diagram illustrating the feature of the user behavior history more specifically. As shown in the figure, for example, features related to the use of shopping at a sales site include features such as the number of purchases, purchase price, purchase date, category of purchased item (for example, fashion, home appliance, food, etc.). The features related to the use of an auction at a sales site include features such as the number of bids, the number of successful bids, and the number of exhibitions. The features related to browsing of the web pages constituting the sales site include features such as a page view, the number of access days of the web page, the latest access date and time, and a search query (search keyword) input at the time of search. For example, the page view may represent the number of accesses from the search site to the sales site, or may represent the number of accesses from the search site to another website (for example, a news site) different from the sale site. . Further, the page view may represent the number of accesses to a website automatically provided by a web browser or the like without going through a search site. That is, the page view may be an index that distinguishes the number of accesses to the website that the user voluntarily accesses from the number of accesses to the website that is passively accessed.

  FIG. 7 is a diagram illustrating an example of the prediction model information 136. As in the illustrated example, the prediction model information 136 is information in which a prediction model is associated with each point. The prediction model is a pattern recognition model generated by a model generation unit 114 described later.

  FIG. 8 is a diagram illustrating an example of the account information 138. As in the illustrated example, the account information 138 is information in which a user ID, a password, an email address, a point balance, and the like are associated with each other for each user.

  Hereinafter, processing by each component of the control unit 110 will be described with reference to a flowchart. FIG. 9 is a flowchart illustrating an example of a series of processes performed by the control unit 110.

  First, the acquisition unit 112 acquires the purchase history information 132 and the feature information 134 from the terminal device 10 or the server device 20 via the communication unit 102 (S100), and stores the acquired information in the storage unit 130. . For example, the acquisition unit 112 acquires the feature indicating the user attribute from the features included in the feature information 134 from the terminal device 10, acquires the feature indicating the user action history from the server device 20, and acquires the feature related to the point. The features acquired from the terminal device 10 and the server device 20 and the features acquired from each device may be combined into user information for each user ID.

  Next, the model generation unit 114 generates a prediction model that models the purchase tendency of the user for each point based on the purchase history information 132 and the feature information 134 acquired by the acquisition unit 112 (S102). The user's purchase tendency is, for example, information indicating the presence / absence of purchase or the amount of purchase required at the time of purchase.

  For example, the model generation unit 114 uses a SVM (Support Vector Machine) to include a feature vector whose element is a result of processing a part or all of a plurality of features included in the feature information 134 or a combination thereof. In the feature space, a hyperplane (a space having a dimension reduced by one from the number of dimensions of the feature space) for deriving each feature vector into a positive example or a negative example is derived. Note that the model generation unit 114 may perform a normalization process on the features to align the order of the values. For example, the positive example is a feature vector indicating a feature of a user who purchased an item after giving points, and the negative example is a feature vector showing a feature of a user who did not purchase an item after giving points. A function representing a hyperplane in the feature space is an example of a prediction model.

  For example, about half of the user features included in the feature information 134 are treated as teacher data (correct data) in SVM, and the remaining user features are treated as learning data. The teacher data is data for deriving the hyperplane described above, and the learning data is data to be classified as one of positive examples and negative examples by the hyperplane derived from the teacher data.

  For example, the model generation unit 114 derives a hyperplane for separating one of positive examples and negative examples from the user's feature at each point, so that each point such as 500 points, 300 points, and 100 points is obtained. The prediction model (function indicating the hyperplane) is generated.

  At this time, the model generation unit 114 derives the hyperplane so that the ratio of the positive example and the negative example of the learning data is approximately the same as the ratio of the positive example and the negative example of the teacher data (preferably the same ratio).

  Then, the model generation unit 114 classifies the feature of the user as the learning data into one of positive examples and negative examples according to the prediction model at each point generated using the teacher data, and derives the score of each classified feature (S104). The score here is information obtained by quantifying the distance between the hyperplane derived as the prediction model and the feature vector indicating the feature.

  FIG. 10 is a diagram illustrating an example of scores derived for each point. As in the illustrated example, a score is derived for each user point. For example, the score “0.465283” at 500 points of the user ID “ABC01” is derived as a prediction model of 500 points with each of a plurality of feature vectors obtained from the user features corresponding to the user ID “ABC01”. It is derived as an average value of the distance to the hyperplane.

  Note that the model generation unit 114 may derive a score at each point for each user by using a learning method such as logistic regression instead of the SVM. In this case, the score is a Wald statistic ((standard error / regression coefficient) ^ 2) or a prediction probability (corresponding to a so-called tendency score) obtained from the Wald statistic.

  Moreover, although the model production | generation part 114 shall score learning data by solving a binary classification problem (problem classified into either a positive example or a negative example) using SVM, it is not restricted to this. . For example, the model generation unit 114 may derive the score at each point by modeling the purchase price for each user's point into a prediction model with reference to the purchase history information 132 shown in FIG. . More specifically, the model generation unit 114 treats the purchase amount by the user for each point included in the purchase history information 132 illustrated in FIG. 4 as teacher data, performs multiple regression analysis, and determines the purchase amount of the teacher data as a model. Turn into. And the model production | generation part 114 derives the score which evaluated the purchase price of learning data by approximating the purchase price of learning data to the prediction model derived | led-out from the teacher data by multiple regression analysis. In this case, the score is a purchase amount determined by multiple regression analysis.

  Next, the purchase probability prediction unit 116 predicts the purchase probability of the user for optimizing the number of purchasers or the purchase price, which is assumed when future points are given based on the purchase results of the past user. (S106). Optimum means, for example, that the number of purchasers or the purchase amount is maximized or near the maximum (for example, about 90% of the maximum value). Hereinafter, in order to simplify the description, it is assumed that optimum = maximum. The purchase probability of the user is an example of “occurrence probability of purchase behavior”.

  For example, the purchase probability prediction unit 116 may represent the future purchase tendency of the user based on the score for each user point derived by the model generation unit 114 and the ratio of positive examples and negative examples of the teacher data. Predict density function. This probability density function is a function that returns a user's future purchase probability with the score for each user's point as a variable. For example, the predicted probability density function is an odd function whose probability approaches 1 as the score increases and whose probability approaches 0 as the score decreases. Such a function is, for example, a sigmoid function, a step function, a linear function, an arctangent function, or other polynomial functions. Hereinafter, as an example, a probability density function representing a user's future purchase tendency will be described as being a sigmoid function represented by Equation (1).

  The SVMScore of Expression (1) represents the score for each user point derived by the model generation unit 114. The parameter A represents the magnitude of the degree of increase or decrease in the probability P that varies according to the SVMScore, and the parameter B represents a component that translates (shifts) the sigmoid function in the SVMScore axis direction.

  Hereinafter, a method for deriving a probability density function representing a user's future purchase tendency will be described. FIG. 11 is a diagram illustrating information regarding purchase histories of all users who have used the sales site in a certain period. The acquisition rate [%] (= A) shown in the figure is an index representing how much the given points have been acquired by the user's intention. For example, the acquisition rate indicates the ratio of the number of users who have received points according to the guidance among the users who are guided to be given points by e-mail, direct mail, push notification, or the like. That is, the acquisition rate indicates the ratio of the number of users who actually received points out of the users counted as point grant candidates. The purchase rate [%] (= B) represents the ratio of the number of users who have purchased items on the sales site during the period until the points expire, out of all users who have acquired points.

  For example, the purchase probability prediction unit 116 determines the parameter A in Equation (1) so that the average value of purchase probabilities returned by the sigmoid function is equal to the purchase rate (= B) in FIG. 11 described above.

  FIG. 12 is a diagram schematically illustrating a method for determining the parameter A. As illustrated (a) to (c), the purchase probability prediction unit 116 changes the magnitude of the parameter A, and the average value of the purchase probabilities returned by the sigmoid function is equal to the purchase rate (= B). Search for parameter A. As shown in the figure, searching for the parameter A means scaling the sigmoid function in the SVMScore axis direction. That is, searching for the parameter A means changing the degree of dispersion of the sigmoid function.

  Note that the parameter B in Equation (1) is suitable for the model generation unit 114 when the ratio of the positive example and the negative example of the learning data in the SVM and the ratio of the positive example and the negative example of the teacher data are approximately the same. It is determined to a correct value.

  FIG. 13 is a diagram schematically illustrating a method for determining the parameter B. For example, as shown in the figure, in the learning stage by SVM, when the ratio of the positive example and the negative example of the teacher data is 15:85, the model generation unit 114 sets the ratio of the positive example and the negative example of the learning data to 15 : Parameter B is set to be close to 85. That is, determining the parameter B means adjusting the numbers of positive examples and negative examples while moving the position of the hyperplane (HP in the figure) in the feature space.

  In this way, the purchase probability prediction unit 116 predicts a probability density function such as a sigmoid function, and substitutes the score derived by the model generation unit 114 for this predicted function, so that the user's future purchase Predict the probability.

  FIG. 14 is a diagram illustrating how the future purchase probability is derived from the score. As shown in the figure, for each user (user ID), the purchase probability prediction unit 116 converts the score of each point (in the example shown, the score by SVM) into a purchase probability according to a probability density function such as a sigmoid function. . In general, when a score obtained by simply applying a learning method such as SVM or logistic regression is used as a probability, the probability density distribution indicating the probability may tend to be a collection of solutions locally. Are known. In contrast, in the present embodiment, a probability density function such as a sigmoid function used to represent an actual event in a different field such as a neural network is not used as a purchase probability based on a score of a prediction model. By fitting the score to the above, it is possible to obtain a probability distribution that matches the actual event even in the field of e-commerce. For example, by searching the parameter A described above so that the average purchase probability returned by the sigmoid function is equal to the actual past purchase rate, the sigmoid function expands and contracts in the SVMScore axis direction, so that the solution is locally Even if they gather together, the probability distribution can be distributed according to the actual event. As a result, the future purchase probability can be accurately predicted.

  Now, the description returns to the flowchart of FIG. Next, the grant point determination unit 116 maximizes the number of purchasers or the purchase price within the budget that can be awarded points based on the purchase probability at each point for each user predicted by the purchase probability prediction unit 116. The distribution of points to be allocated to each user is determined (S108).

  For example, the grant point determination unit 116 uses integer programming to determine the allocation of points to be allocated to each user so as to maximize the number of purchasers or the purchase price within a budget in which points can be granted. The integer programming in the present embodiment refers to a high-dimensional vector (hereinafter referred to as a point distribution vector V) having vector components for point types n (in this embodiment, n = 5) as elements e. By multiplying the purchase probability for each point of the user, a process for finding a solution (a combination of points to be allocated to the user) that satisfies the constraints and has the optimal objective function. The element e of the point distribution vector V takes one of the values 1 or 0, and the sum of all the elements is 1. In other words, only one of the elements e of the point distribution vector V takes a value of 1 and the values of the other elements are 0.

  The objective function is the sum of purchase probabilities of all users, and “optimize the objective function” means that the sum of purchase probabilities of all users is maximized by giving points. The objective function may be the sum of purchase amounts of all users. In this case, “optimize the objective function” means that the sum of purchase amounts of all users is maximized by giving points. It may be to.

The constraint conditions include, for example, the following conditions.
(1) Keeping the total amount of points to be granted within the budget of points that can be granted (budget ≦ total amount of grant points).
(2) The element e of the point allocation vector V is an integer of either 1 or 0.
(3) For the element e of the point distribution vector V, the sum of all elements is 1.

  When the processing capability of the processor is limited, the grant point determination unit 116 performs a process for obtaining an optimal solution of the objective function for a part of the data of the total number of users, It may be handled as a data processing result. Some data is data including the number of samples (number of users) to a degree that is statistically sufficiently reliable, for example. Specifically, if the total number of users is about 10 million, the number of users of about 500,000 may be included in some data.

  Further, the grant point determination unit 116 may further obtain an optimal solution using a linear programming relaxation technique as one technique of integer programming. The linear program relaxation in the present embodiment is to search for an optimal solution for each point while treating the element e of the above-described point allocation vector V as a decimal number (for example, 0.5) that can be taken between 0 and 1, In the final calculation process, add a condition that the solution is an integer, convert the decimal variable (eg purchase probability x element value) to an integer, and choose the largest integer among the converted integers It is a technique to obtain as a solution. At this time, even if each element is a decimal, the sum of the elements is 1.

  FIG. 15 is a diagram for explaining linear program relaxation. As shown in the figure, for example, when a total of five types of points of 500, 300, 100, 3, and 0 are given, the grant point determination unit 116 obtains the purchase probability for each of the five types of points for each user ID. When the optimization calculation is performed and the variable (purchase probability x element value) obtained in the optimization process becomes extremely small or the element value becomes small enough to approximate 0, the point is The optimization calculation is performed again at the remaining points without consideration.

  More specifically, the value of the vector element e1 that multiplies the purchase probability P1 of 500 points is 0.30, the value of the vector element e2 that multiplies the purchase probability P2 of 300 points is 0.20, and the purchase probability P3 of 100 points The value of the vector element e3 to be multiplied by 0.29 is 0.29, the value of the vector element e4 to be multiplied by the 3-point purchase probability P4 is 0.01, and the value of the vector element e5 by which the 0-point purchase probability P5 is multiplied is 0.20. In this case, the granted point determination unit 116 performs optimization calculation again with the remaining 500 points, 300 points, 100 points, and 0 points without considering giving 3 points. Thereby, the distribution ratio of the vector element e to the remaining points can be changed. If the value of the vector element e allocated to any point is 0, the recalculation may be omitted. In the example of FIG. 15, any vector element e may be zero. In this case, the grant point determination unit 116 may exclude the target point to be multiplied by the vector element e from the optimization calculation target.

  In the example described above, a total of five types of points 500, 300, 100, 3, and 0 have been prepared. However, the present invention is not limited to this. For example, other types of points (points that have not been given in the past) such as 400 points or 200 points may be added, or any of the above may be omitted.

  Further, purchase probabilities of points that have not been given in the past (for example, 400 points, 200 points, etc.) may be complemented by performing linear interpolation, linear approximation complement, curve approximation complement, or the like. For example, when the grant point determination unit 116 estimates the purchase probability of 200 points that have not been granted by performing linear interpolation, the average of the purchase probability of 300 points and the purchase probability of 100 points is calculated as 200 points that have not been granted. May be derived as the purchase probability.

  As a result, the granted point determination unit 116 maximizes the number of purchasers and the purchase price by combining various points such as 450, 350, 250, 50, 3, and 0 points as points that have not been granted in the past. It is possible to search for an optimal combination of points by repeatedly performing the above process. As in the example described above, even when there are only 500, 300, 100, 3, 0 point history as points that have been actually given, these points were derived according to the respective prediction models. Using five purchase probabilities, points that have not been granted in the past can be obtained by linear interpolation or the like. For example, the grant point determination unit 116 treats all natural numbers (0, 1, 2, 3,..., 499, 500) that can be taken in a range with 0 point as the lower limit and 500 points as the upper limit, as points, Purchase probabilities corresponding to these 501 types of points are obtained by linear interpolation or the like. Then, for each user ID, the grant point determination unit 116 selects a point that maximizes the number of purchasers and the purchase amount from among the 501 types of points, and determines to grant this. In this case, the number of elements e of the element e of the point distribution vector V is 501, and only one of the values of the elements is 1. Note that the number of points to be compensated by complementation is not limited to 501 and may be 101 types (0, 5, 10,..., 495, 500) in 5 increments in the range from 0 to 500, for example.

  Further, the given point determination unit 116 uses the least square method or the like to calculate the purchase probability of each point of 500, 300, 100, 3, 0, and return the purchase probability using the point value as a variable. A function may be derived. In this case, the grant point determination unit 116 may obtain a purchase probability corresponding to the unassigned point by substituting the unassigned point value into the derived function.

  In this way, it is possible to estimate a purchase probability corresponding to a point that has not been given in the past by linear interpolation or linear approximation complementation. As a result, when new types of points (points with different amounts) that have not been granted in the past are to be newly granted this time, the number of purchasers or purchase amount assumed by giving new types of points to the user It is possible to derive a suitable combination of points to be given so that is maximized.

  In addition to the linear program relaxation method, the grant point determination unit 116 may use, for example, a Lagrangian relaxation method, or a heuristic such as an annealing method, a tabu search method, a random multi-start local optimization method, or a genetic algorithm. By using the text method, the distribution of points allocated to each user may be determined so as to maximize the number of purchasers or the purchase amount.

  Next, the grant point determination unit 116 notifies the user of guidance indicating that the points determined using the integer programming have been granted via the communication unit 102 (S110). For example, the grant point determination unit 116 refers to the account information 138 and sends an e-mail to a mail address associated with the user ID, thereby notifying the user of point grant guidance. In this mail, for example, a URL (Uniform Resource Locator) to a website capable of receiving points is described. This URL-destination website may be, for example, a website where login is performed by inputting a user ID used at a sales site. For example, when the user operates the terminal device 10 to access a predetermined website from this URL, points assigned to the user ID input at the time of login are given. Thereby, in the account information 138, the given points are added to the point balance corresponding to the user ID. The grant point determination unit 116 may notify the terminal device 10 that can be operated by the user by a push notification or the like instead of the e-mail, or send a direct mail to an address set in advance on the sales site by the user. The guidance may be notified by mail. The grant point determination unit 116 may grant the optimal point determined for each user regardless of the user's intention to receive the information, and notify the user by e-mail or the like afterwards. .

  FIG. 16 is a diagram illustrating an example of optimal points allocated for each user. In the illustrated example, the point budget is set to less than 2000, and points are allocated to each user so that the purchase probability is highest within this budget. For example, when the total purchase probability when 300 points are uniformly given to all users and the total purchase probability when optimum points are given for each user in a range not exceeding the point budget, The total purchase probability is higher when the optimal point is given than when the point is given. As a result, the number of purchasers can be increased within a limited point budget.

  In addition, in practice, the optimal points obtained in the present embodiment are given to the number of purchasers when 300 points are uniformly given to the target users and the same number of users as those who are given 300 points. When comparing with the number of purchasers in the case of granting, the number of purchasers was improved by about 32% in the case of giving optimum points to individual users as compared to the case of giving points uniformly. Note that these points are given at the same time.

  According to the first embodiment described above, a purchase history of a user to which any one or more incentives (for example, points) having different degrees of influence on the user are given, and incentives are given. The acquisition unit 112 that acquires the feature information for each user, and the model generation unit 114 that generates a prediction model that models the purchase tendency by the user for each incentive based on the purchase history of the user and the feature information for each user. And predicting a probability for optimizing an event related to purchase behavior (for example, the number of purchasers and the purchase price) caused by giving at least one incentive out of a plurality of incentives based on a prediction model. By providing the purchase probability prediction unit 116, it is possible to more appropriately relate to purchase behavior. It is possible to determine the rate.

  In addition, according to the first embodiment described above, incentives are effectively given because points are given to each user so as to maximize the number of purchasers or the purchase price within a limited point budget. can do.

  In addition, according to the first embodiment described above, when the prediction model is generated, in addition to the feature related to the user attribute and the feature related to the user behavior, the feature related to the point is further used. If the user to be used has a large purchase frequency and the amount of money required for the purchase, it is possible to carry out a point allocation method that gives more points to such a user.

  In addition, when using features related to points at the time of generating a prediction model, it is possible to generate a prediction model that reflects the correlation of user attributes and the correlation of user behavior with respect to the usage tendency of points. It is possible to analyze a user's purchase tendency in various ways.

  In addition, according to the first embodiment described above, whether the user voluntarily accesses the website or passively accesses the web page browsing feature as the feature related to the user's behavior. The purchase tendency of the user can be analyzed including the above. By this, for example, if the number of purchases is larger as the user voluntarily accesses the sales site, the number of purchasers can be more effectively within the point budget by preferentially allocating points to other users other than the user. The purchase price can be increased.

<Second Embodiment>
Hereinafter, the second embodiment will be described. The information analysis apparatus 100 according to the second embodiment is different from the first embodiment described above in that the purchase probability is changed according to the point acquisition rate and usage rate for each user. The following description will focus on differences from the first embodiment, and description of points that are the same as in the first embodiment will be omitted. In the description of the second embodiment, the same parts as those in the first embodiment will be described with the same reference numerals.

  When the point acquisition rate is included in the purchase history information 132 for each user, the grant point determination unit 116 according to the second embodiment points based on the point acquisition rate and the purchase probability at each point. Within the budget that can be assigned, the distribution of points allocated to each user is determined so as to maximize the number of purchasers or the purchase amount. The point acquisition rate in the second embodiment is an index that represents how much points are acquired by the user's intention for each user.

  For example, as described above, the user may be notified by email that points have been granted. The email includes, for example, the URL of a website that requires login with a user ID, information for notifying the user that points are given by login, and the like. When the user operates the terminal device 10 to access the website of the URL destination, the granted point determination unit 116 determines that the user has an intention to receive the point, and determines the allocation to the user ID input at the time of login. Points are given to the user by associating these points.

  On the other hand, when there is no login from the URL even though the mail is transmitted, the grant point determination unit 116 determines that the user does not intend to receive points, and does not grant points to the user.

  In this way, the granted point determination unit 116 does not subtract points that were not actually acquired from the point budget in consideration of the acquisition rate (the number of accesses from the URL / the total number of email transmissions) for each user. You may assign points to For example, in the course of the point allocation process, the grant point determination unit 116 uses a product obtained by multiplying the purchase probability converted from the score by a fitting function such as a sigmoid function by the acquisition rate for each point in integer programming. It may be the final purchase probability handled. As a result, the total amount of points easily falls below the point budget, and the number of purchasers and the purchase amount can be increased. When points are awarded regardless of the user's intention to receive, the point acquisition rate may be treated as 100%.

  Further, the grant point determination unit 116 predicts when the purchase history information 132 for each user does not include the point acquisition rate and includes information such as “has acquired points” and “has not acquired points”. Similar to the binary classification problem of “purchased” and “not purchased” using the model, these may be regarded as a binary classification problem and the point acquisition rate may be derived. For example, the grant point determination unit 116 may derive a score using SVM, logistic regression, or the like, and derive an acquisition rate from the score using a fitting function such as the sigmoid function described above.

  Further, when the point usage rate is included in the purchase history information 132 for each user, the grant point determination unit 116 can give points based on the point usage rate and the purchase probability at each point. Within the budget, the distribution of points allocated to each user may be determined so as to maximize the number of purchasers or the purchase amount. The point usage rate is an index representing how much points are used within a predetermined period. For example, the point utilization rate decreases as the number of users who have not used the points after the points are given and before the predetermined period elapses. Therefore, the grant point determination unit 116 may use a product obtained by multiplying the purchase probability converted from the score by the fitting function by the utilization rate for each point as the final purchase probability handled in the integer programming. As a result, points that are assumed not to be used among points to be given can be considered in advance, so that more points are given to a user who is likely to cause purchase behavior for point granting. be able to. As a result, points can be given more effectively, and the number of purchasers and the purchase amount can be further increased.

  Further, the grant point determination unit 116 multiplies the purchase probability converted from the score by the fitting function by both the acquisition rate and the utilization rate for each point as the final purchase probability handled in the integer programming method. Also good.

  In addition, when the purchase history information 132 for each user does not include the point usage rate and includes information such as “uses points” and “does not use points”, the grant point determination unit 116 Similarly to the method of deriving the acquisition rate, these may be regarded as a binary classification problem and the point utilization rate may be derived.

  Further, the granted point determination unit 116 may derive the point usage rate based on the point acquisition rate. For example, since it is known that there is a linear relationship between the point acquisition rate and the usage rate, the grant point determination unit 116 uses a point multiplication rate obtained by multiplying the point acquisition rate by a constant. It may be derived as a rate.

  According to the second embodiment described above, as in the first embodiment described above, a prediction model in which the purchase tendency by the user for each incentive is modeled based on the purchase history of the user and the feature information for each user. Probability for optimizing events related to purchase behavior (for example, the number of purchasers and purchase price) that occur by giving at least one incentive out of multiple incentives based on this prediction model By predicting, the probability relating to purchase behavior can be obtained more appropriately.

<Other embodiments>
Hereinafter, as other embodiments, modifications of the above-described embodiment will be described. In the above-described embodiment, either the sum of purchase probabilities of all users who are scheduled to give points or the sum of purchase amounts of all users is treated as an objective function and given to the user to maximize either of them. However, the present invention is not limited to this. For example, both the number of purchasers and the purchase price may be used as the objective function. In this case, for example, the grant point determination unit 116 solves the problem of the integer programming method using the linear sum of the number of purchasers and the purchase amount as an objective function, and thereby, within the budget that can give points, The distribution of points to be allocated to each user is determined so as to maximize the objective function considering both purchase amounts.

  In the above-described embodiment, the purchase probability, the acquisition rate, and the usage rate are described as different for each user's point, but the present invention is not limited thereto. For example, some or all of the purchase probability, acquisition rate, and usage rate may be common to all users. For example, as shown in FIG. 11 described above, the grant point determination unit 116 may use the purchase probability of all users averaged for each point as the purchase probability of each user.

  In the above-described embodiment, the points to be allocated to the user are determined by performing integer programming, but the present invention is not limited to this. For example, the points to be allocated to the user may be determined using the heuristic method described above. For example, the given point determination unit 116 assumes a certain arbitrary point as a provisional optimum solution (approximate solution), and searches for a point where the objective function becomes larger while satisfying the constraint condition when the optimum solution is changed. To do. This method is a so-called local search. For example, the provision point determination unit 116 assumes 3 points as a provisional optimal solution, and if the value of the objective function when changing from 3 points to 100 points is larger than the value of the objective function at 3 points, the provisional point determination unit 116 The optimal solution is changed from 3 points to 100 points. Furthermore, the grant point determination unit 116 searches for a target function having a larger value while sequentially changing target points such as 300 points and 500 points. As a result, the assigned point determination unit 116 derives the point having the largest objective function value as the optimum solution. Note that the above-described heuristic technique is merely an example, and other techniques may be used.

<Hardware configuration>
Of the plurality of devices included in the information analysis system 1 of the above-described embodiment, at least the information analysis device 100 is realized by a hardware configuration as illustrated in FIG. 17, for example. FIG. 17 is a diagram illustrating an example of a hardware configuration of the information analysis apparatus 100 according to the embodiment.

  The information analyzing apparatus 100 includes an NIC 100-1, a CPU 100-2, a RAM 100-3, a ROM 100-4, a secondary storage device 100-5 such as a flash memory and an HDD, and a drive device 100-6, which are connected to an internal bus or a dedicated communication line. Are connected to each other. The drive device 100-6 is loaded with a portable storage medium such as an optical disk. A program stored in a portable storage medium attached to the secondary storage device 100-5 or the drive device 100-6 is expanded in the RAM 100-3 by a DMA controller (not shown) or the like and executed by the CPU 100-2. Thus, the control unit 110 is realized. The program referred to by the control unit 110 may be downloaded from another device via the network NW.

  As mentioned above, although the form for implementing this invention was demonstrated using embodiment, this invention is not limited to such embodiment at all, In the range which does not deviate from the summary of this invention, various deformation | transformation and substitution Can be added.

DESCRIPTION OF SYMBOLS 1 ... Information analysis system, 10 ... Terminal device, 20 ... Server apparatus, 100 ... Information analysis apparatus,
DESCRIPTION OF SYMBOLS 102 ... Communication part, 110 ... Control part, 112 ... Acquisition part, 114 ... Model production | generation part, 116 ... Purchase probability prediction part, 118 ... Grant point determination part, 130 ... Memory | storage part, 132 ... Purchase history information, 134 ... Feature information 136 ... Prediction model information, 138 ... Account information, NW ... Network

Claims (11)

An acquisition unit that acquires a purchase history of a user to which any one or more incentives are given among a plurality of incentives having different degrees of influence on the user, and feature information for each user to which the incentives are given,
Based on the purchase history of the user acquired by the acquisition unit and the feature information for each user, a model generation unit that generates a prediction model that models the purchase tendency by the user for each incentive;
Based on the prediction model generated by the model generation unit, a probability prediction unit that predicts a probability for optimizing an event related to purchase behavior caused by giving at least one incentive among the plurality of incentives When,
An information analysis apparatus comprising: The probability prediction unit predicts a probability for optimizing the number of users who cause purchase behavior by giving the incentive,
The information analysis apparatus according to claim 1. The model generation unit generates the prediction model by a learning model in which teacher data is the presence or absence of purchase by the user for each incentive,
The information analysis apparatus according to claim 1 or 2. The probability prediction unit predicts a probability for optimizing the amount of money to be purchased by giving the incentive,
The information analysis apparatus according to any one of claims 1 to 3. The probability prediction unit generates the prediction model by a learning model using the amount of purchase by the user for each incentive as teacher data.
The information analysis device according to any one of claims 1 to 4. The probability prediction unit approaches the score of the prediction model generated by the model generation unit as the probability increases to 1 as the score increases and the probability approaches 0 as the score decreases. Based on the function, the event related to the purchase behavior is converted into a probability for optimizing,
The information analysis apparatus according to any one of claims 1 to 5. The feature information for each user includes some or all of user attributes, user behavior history, and information on incentives given to users in the past.
The information analysis apparatus according to any one of claims 1 to 6. The probability prediction unit, based on the prediction model corresponding to the incentive that has been given to the user at a past time point, out of the plurality of incentives, the incentive that has not been given to the user at a past time point Predict the probability of optimizing events related to purchase behavior that can occur by granting to users at a future time,
The information analysis apparatus according to any one of claims 1 to 7. An acquisition unit that acquires a purchase history of a user to which any one or more incentives are given among a plurality of incentives having different degrees of influence on the user, and feature information for each user to which the incentives are given,
Based on the purchase history of the user acquired by the acquisition unit and the feature information for each user, a model generation unit that generates a prediction model that models the purchase tendency by the user for each incentive;
Based on an odd function, the probability of the prediction model generated by the model generation unit approaches 1 as the score increases and the probability approaches 0 as the score decreases. A probability prediction unit that converts a probability for optimizing an event related to purchase behavior caused by giving at least one incentive out of the plurality of incentives;
An information analysis apparatus comprising: Computer
Obtaining a purchase history of a user to which any one or more incentives are given among a plurality of incentives having different degrees of influence on the user, and feature information for each user to which the incentives are given,
Based on the acquired purchase history of the user and the feature information for each user, generate a prediction model that models the purchase tendency by the user for each incentive,
Based on the generated prediction model, predicting a probability for optimizing an event related to purchase behavior caused by giving at least one incentive out of the plurality of incentives,
Information analysis method. On the computer,
Among a plurality of incentives having different degrees of influence on the user, the purchase history of the user given any one or more incentives and the feature information for each user given the incentives are acquired,
Based on the acquired purchase history of the user and the feature information for each user, to generate a prediction model that models the purchase tendency by the user for each incentive,
Based on the generated prediction model, a probability for optimizing an event related to purchase behavior caused by giving at least one incentive out of the plurality of incentives is predicted.
Information analysis program.

Images