JP2012247860A - Model generation method and model generation device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To measure loyalty of a user whose purchase history is not available.SOLUTION: A model generation device generates a plurality of alternative models having explanation variables replaced with other history information other than purchase histories for an existent model whose target variable as loyalty is found based upon a purchase history of articles as an explanation variable. Then the model generation device calculates evaluated values indicative of adaptation of other history information to the alternative models. Further, the model generation device receives input of a usable variable as a usable explanation variable. Furthermore, the model generation device extracts alternative models having other history information included in usable variables from the plurality of alternative models, and presents an alternative model having the highest evaluated value among the extracted alternative models.

Description

  Embodiments described herein relate generally to a model generation method and a model generation apparatus.

  Traditionally, in the marketing field, customer loyalty has been measured in order to take strategies such as direct mail (DM) transmission only for customers with high interest in corporate brands. It has been broken. “Royalty” as used herein indicates an awareness of the customer (user) with respect to the corporate brand and is an index indicating whether the customer will continue to purchase products in the future. For example, a customer with a high loyalty to a corporate brand is likely to purchase a product of the corporate brand.

  In general, such loyalty is often measured by a technique called “RFM analysis”. In this RFM analysis, it can be said that the royalty, which is the objective variable, is measured using R (the most recent purchase date), F (purchase frequency), and M (cumulative purchase price) as explanatory variables. For example, when a company provides a shopping site, the loyalty of each user is R (recent purchase date), F (purchase frequency), M (of the purchase history accumulated by the user on the shopping site). Measured from explanatory variables such as (total purchase price).

INTERNET MARKETING NEWS, “RFM Analysis (1)”, [online], [Search May 18, 2011], Internet <http://www.newsbit.info/note/20010912.html>

  However, it is difficult to measure the loyalty of a user who has no purchase history in the above-described conventional technology. Specifically, the royalty measurement method in the conventional RFM analysis, etc., as described above, measures the royalty using the purchase history of R, F, M, etc., so the royalties of users who have never purchased the product It is difficult to measure the tee.

  This invention is made in view of the above, Comprising: It aims at providing the model production | generation method and model production | generation apparatus which make it possible to measure a user's royalty without a purchase history.

  The model generation method according to the embodiment is a model generation method executed by a model generation device, and an existing model that is a model for obtaining an objective variable that is a user's loyalty using a purchase history of a product by a user as an explanatory variable. On the other hand, a model generation step of generating a plurality of alternative models, which are models obtained by replacing the explanatory variables with other history information other than the purchase history, and for each alternative model generated by the model generation step, An evaluation value calculating step for calculating an evaluation value indicating the degree of fitness of the other history information, a receiving step for receiving an input of an available variable that is a usable explanatory variable, and a plurality of alternatives generated by the model generating step Extract an alternative model that includes the other history information in the usable variables accepted by the acceptance step from the model, and extract Evaluation value calculated by the evaluation value calculating step of alternative models is characterized by including a model presentation step of presenting the highest alternative model.

  The model generation method and the model generation apparatus according to the embodiment have an effect that it is possible to measure a user's loyalty without a purchase history.

FIG. 1 is a diagram illustrating an example of processing performed by the model generation apparatus according to the first embodiment. FIG. 2 is a block diagram illustrating a configuration example of the model generation apparatus according to the first embodiment. FIG. 3 is a diagram illustrating an example of learning data according to the first embodiment. FIG. 4 is a diagram illustrating an example of an alternative model group in the first embodiment. FIG. 5 is a diagram illustrating an example of processing performed by the model presenting unit according to the first embodiment. FIG. 6 is a flowchart illustrating a model generation processing procedure performed by the model generation apparatus according to the first embodiment. FIG. 7 is a flowchart of a model generation process procedure performed by the model generation apparatus according to the first embodiment. FIG. 8 is a flowchart of the alternative model presentation process procedure performed by the model generation apparatus according to the first embodiment. FIG. 9 is a diagram illustrating a computer that executes a model generation program.

  Embodiments of a model generation method and a model generation apparatus according to the present invention will be described below in detail with reference to the drawings. In addition, this invention is not limited by this Example. In the following embodiment, an example of measuring the loyalty of each user who uses a shopping site that sells products via the Internet will be described. The “model” described in the following embodiments indicates a function (calculation formula) for measuring loyalty that is an objective variable.

[Processing example of model generator]
First, processing by the model generation apparatus according to the first embodiment will be described with reference to FIG. FIG. 1 is a diagram illustrating an example of processing performed by the model generation apparatus according to the first embodiment. In the example shown in FIG. 1, the analyst P1 measures the loyalty of the user who uses the shopping site with respect to the company brand.

  Here, a user who uses the shopping site browses a product introduction page or purchases a product by accessing the shopping site. Among such users, there are users who have browsed and purchased products, and there are users who have browsed products but have not purchased products. The analyst P1 shown in FIG. 1 also sets a royalty measurement target for a user who has not purchased a product in the past on a shopping site. That is, an existing model used in RFM analysis or the like (hereinafter may be referred to as “existing model”) is a purchase history of the user (R: most recent purchase date, F: purchase frequency, M: cumulative purchase amount, etc.) Therefore, it is difficult for the analyst P1 to measure the loyalty of each user using the existing model.

  Therefore, the model generation apparatus 100 according to the first embodiment presents the analyst P1 with a model for measuring user loyalty without a purchase history (hereinafter, sometimes referred to as “alternative model”). , Make it possible to measure the loyalty of each user.

  Briefly explaining this point, the analyst P1 uses the information processing apparatus 1 to provide the model generation apparatus 100 with explanatory variables that can be used (hereinafter may be referred to as “usable variables”). Send. For example, the analyst P1 transmits the browsing history (the most recent browsing date, browsing frequency, etc.) to the model generation device 100 when the user who is the target of royalty measurement has only browsed the product. In such a case, the model generation device 100 according to the first embodiment transmits an alternative model having the usable variable received from the information processing device 1 as an explanatory variable to the information processing device 1. Thus, the analyst P1 can measure the loyalty of the user who has no purchase history using the alternative model provided from the model generation device 100.

  Such processing by the model generation apparatus 100 will be described with reference to FIG. As shown in FIG. 1, the model generation apparatus 100 according to the first embodiment holds learning data 10a and correct answer data 10b. The learning data 10a is history information including purchase information related to purchase of products by each user who uses the shopping site, browsing information related to product browsing, and the like. The correct answer data 10b is purchase presence / absence information indicating whether each user has actually purchased a product.

  In the first embodiment, the learning data 10a is past information than the correct answer data 10b. For example, when the learning data 10a is purchase information or browsing information in April, May, and June 2010, the correct answer data 10b is purchase presence / absence information in July 2010. The reason why the model generation apparatus 100 holds such learning data 10a and correct answer data 10b will be described below.

  The model generation device 100 calculates an evaluation value of the existing model 21 using the purchase information and the correct answer data 10b included in the learning data 10a. The “evaluation value” here refers to the degree of fitness (“applicable”) of the purchase information (R: data of latest purchase date, F: purchase frequency, M: cumulative purchase price, etc.) included in the learning data 10a. For example, the AIC (Akaike Information Criterion). The “existing model” in the first embodiment is an existing analysis model used in RFM analysis or the like as described above, and purchase information (R: latest purchase date, F: purchase frequency, M: cumulative purchase amount, etc.) Data) is assumed to be represented by a predetermined function having explanatory variables.

  Specifically, the model generation device 100 measures the loyalty of each user by applying the purchase information of the learning data 10a to the existing model 21. Then, the model generation device 100 compares the measured royalty with the correct answer data 10b. Thereby, the model generation apparatus 100 can determine how much the measured royalty is correct, and calculates the correctness of this royalty as an evaluation value. If it demonstrates using the said example, the model production | generation apparatus 100 will measure the loyalty of each user by applying the learning data 10a (purchase information) in April, May, and 2010 to the existing model 21, for example. To do. Then, in order to evaluate the accuracy of the measured loyalty, the model generation apparatus 100 determines whether or not each user has actually purchased a product using the correct answer data 10b in July 2010. That is, the model generation device 100 uses the learning data 10a to measure loyalty, and uses the correct answer data 10b, which is future information, rather than the learning data 10a in order to determine the accuracy of the measured loyalty.

  In the case of a user who has purchased a product in the past, purchase information is included in the learning data 10a, but the product has not been purchased in the past and has only been viewed. In the case of a user, purchase information is not included in the learning data 10a. Therefore, the model generation device 100 calculates the evaluation value of the existing model 21 using purchase information of a user who has purchased a product in the past.

  Subsequently, the model generation apparatus 100 according to the first embodiment uses the browsing information included in the learning data 10a and the like, except for the explanatory variable R (the most recent purchase date), F (the purchase frequency), or M (the accumulated purchase price). An alternative model group 30 represented by the explanatory variables is generated. Specifically, the model generation device 100 replaces the explanatory variables R, F, M, and the like included in the existing model 21 with explanatory variables such as browsing information (the latest browsing date, browsing frequency, etc.). Is generated. Here, since there are a plurality of patterns such as browsing information that can be substituted for R, F, or M, the model generation apparatus 100 generates a replacement model for each pattern of browsing information that can be substituted for R, F, or M. By doing so, a plurality of alternative models are generated. Although FIG. 1 shows an example in which the model generation device 100 generates the alternative models 31 and 32, the model generation device 100 may generate three or more alternative models.

  And the model production | generation apparatus 100 calculates evaluation value (AIC etc.) for every alternative model using the browsing information contained in the learning data 10a, and the correct answer data 10b. At this time, the model generation device 100 calculates the evaluation value of the alternative model group 30 in the same manner as the method of calculating the evaluation value of the existing model 21.

  As described above, the model generation apparatus 100 calculates the evaluation value of the existing model 21 having the purchase information (R, F, and M) as explanatory variables, and generates an alternative model group 30 having the browsing information and the like as explanatory variables. The evaluation value of the alternative model group 30 is calculated.

  Then, when the model generation apparatus 100 accepts usable variables from the information processing apparatus 1, the alternative model group 30 includes all explanatory variables in the usable variables and has the highest evaluation value. To extract. In the example illustrated in FIG. 1, the model generation apparatus 100 extracts an alternative model 31 from the alternative model group 30. Then, the model generation device 100 transmits the alternative model 31 and the evaluation value of the alternative model 31, and the existing model 21 and the evaluation value of the existing model 21 to the information processing device 1.

  Thereby, the analyst P1 can measure the loyalty even for the user who has not purchased the product in the past by applying the usable variable to the alternative model 31. Note that the alternative model 31 generated by the model generation apparatus 100 includes the explanatory variables of purchase information (R: latest purchase date, F: purchase frequency, M: cumulative purchase price) included in the existing model 21 as browsing information or the like. Since it has been replaced, there is a possibility of measuring loyalty with a lower accuracy than the existing model 21. However, the model generation device 100 transmits the evaluation value of the existing model 21 together with the evaluation value of the alternative model 31 to the information processing device 1. For this reason, the analyst P1 can grasp the degree of accuracy with which the alternative model 31 can measure the loyalty compared to the existing model 21.

  Below, the model generation apparatus 100 mentioned above is demonstrated in detail. In the following, a configuration example of the model generation device 100 according to the first embodiment will be described first, then a processing procedure performed by the model generation device 100 according to the first embodiment will be described, and finally, a model according to the first embodiment will be described. The effect by the production | generation apparatus 100 is demonstrated.

[Example of model generator configuration]
A configuration example of the model generation device 100 according to the first embodiment will be described with reference to FIG. FIG. 2 is a block diagram illustrating a configuration example of the model generation device 100 according to the first embodiment. As illustrated in FIG. 2, the model generation device 100 according to the first embodiment includes learning data 10, a model generation unit 110, a reception unit 120, and a model presentation unit 130.

  The learning data 10 is history information including purchase information regarding purchase of products by each user, browsing information regarding browsing of products by each user, and the like. The learning data 10 shown in FIG. 2 includes not only the history information but also the correct answer data 10b shown in FIG. That is, the learning data 10 shown in FIG. 2 is data in which the learning data 10a and the correct answer data 10b shown in FIG. 1 are integrated.

  Here, the learning data 10 will be described with reference to FIG. FIG. 3 is a diagram illustrating an example of the learning data 10 according to the first embodiment. FIG. 3 shows an example of items included in the learning data 10. As illustrated in FIG. 3, the learning data 10 in the first embodiment includes a user identifier “ID”, the number of purchases “x”, the number of purchase days “t”, the number of first purchase days “T”, the purchase survival probability “life”, Browsing count “view_x”, browsing days “view_t”, initial browsing days “view_T”, browsing survival probability “view_life”, cart usage count “cart_x”, cart usage count “cart_t”, initial cart usage count “cart_T”, cart usage The survival probability “cart_life”, the purchase cycle “buy_cycle”, the purchase presence / absence “real_life”, and the like are included.

  The user identifier “ID” is an identifier for identifying a user who uses the shopping site, and may be information such as “ID11”, for example. The learning data 10 stores a user identifier “ID” of a user who has browsed or purchased a product among the users registered in the shopping site.

  The number of times of purchase “x” indicates the number of times the product has been purchased, and for example, information such as “10 times” can be taken. The number of days of purchase “t” indicates the number of days elapsed from the date of first purchase of the product to the date of last purchase of the product, and may be information such as “20 days”. The first purchase days “T” indicates the number of days that have elapsed from the date when the product is first purchased to the current date, and can be information such as “30 days”, for example. The purchase survival probability “life” is loyalty indicating whether or not the user continues to purchase the product, and may be information such as “20%”. This purchase survival probability “life” is an existing model for measuring loyalty related to product purchase, and is a model (function) having the number of purchases “x”, the number of purchases “t”, and the number of first purchases “T” as explanatory variables. ). That is, the purchase survival probability “life” can be expressed as “life (x, t, T)”.

  The number of times of browsing “view_x” indicates the number of times the product is browsed, and information such as “120 times” can be taken. The number of browsing days “view_t” indicates the number of days that have elapsed from the date when the product was first browsed to the date when the product was last browsed, and information such as “10 days” can be taken. The first viewing date “view_T” indicates the number of days elapsed from the date when the product is first browsed to the current date, and may be information such as “20 days”. The browsing survival probability “view_life” indicates whether or not the user continues to browse the product, and can take information such as “50%”, for example. The browsing survival probability “view_life” is represented by a model (function) having the browsing count “view_x”, the browsing days “view_t”, and the initial browsing days “view_T” as explanatory variables. Specifically, the browsing survival probability “view_life” includes “x”, “t”, and “T” of the above “life (x, t, T)” as the number of browsing times “view_x”, the number of browsing days “view_t”, This model is replaced with the first viewing date “view_T”. That is, the browsing survival probability “view_life” can be expressed as “life (view_x, view_t, view_T)”.

  The cart use count “cart_x” indicates the number of times the product has been put into the cart, and information such as “60 times” can be taken. Note that “cart” in the present embodiment indicates a shopping cart used in a shopping site, and has a function for a user to temporarily manage products to be purchased. The cart use days “cart_t” indicates the number of days elapsed from the date when the product is first put into the cart to the date when the product is finally put into the cart. For example, information such as “15 days” can be taken. The first cart use days “cart_T” indicates the number of days elapsed from the date when the product is first put into the cart to the current date, and may be information such as “22 days”. The cart use survival probability “cart_life” indicates whether or not the user continues to put the product into the cart, and may be information such as “40%”, for example. The cart use survival probability “cart_life” is represented by a model having the cart use count “cart_x”, the cart use days “cart_t”, and the first cart use days “cart_T” as explanatory variables. Specifically, the cart use survival probability “cart_life” includes “x”, “t”, and “T” of the “life (x, t, T)” as the cart use count “cart_x” and the cart use days “ “cart_t” and the first cart use days “cart_T”. That is, the cart use survival probability “cart_life” can be expressed as “life (cart_x, cart_t, cart_T)”.

  The purchase cycle “buy_cycle” indicates a cycle in which the user purchases a product, and information such as “10 days / time” can be taken. In the case of this example, it indicates that a product is purchased once every 10 days. The number of purchases “x” to the purchase cycle “buy_cycle” so far corresponds to the learning data 10a shown in FIG.

  The purchase presence / absence “real_life” indicates whether or not the user has purchased the product. In the following, if purchase presence / absence “real_life” = “1”, it indicates that the user has purchased the product, and if purchase purchase / non-existence “real_life” = “0”, the user has purchased the product. It shall be shown that there is not. Such purchase presence / absence “real_life” corresponds to the correct answer data 10b shown in FIG.

  That is, in the learning data 10, the number of purchases “x” to the purchase cycle “buy_cycle” are past information than the purchase presence / absence “real_life”. For example, the number of purchases “x” to the purchase cycle “buy_cycle” is information in April, May, and June 2010, and the purchase presence / absence “real_life” is information in July 2010, for example. The various items illustrated in FIG. 3 are examples, and the learning data 10 may include items other than the information illustrated in FIG.

  Returning to the description of FIG. 2, the model generation unit 110 uses the learning data 10 to calculate an evaluation value of the existing model, further generates a plurality of alternative models, and calculates an evaluation value of the alternative model. . In the following, the processing by the model generation unit 110 will be described separately for (A) an existing model evaluation value calculation process, and (B) an alternative model generation process and an alternative model evaluation value calculation process.

  (A) The evaluation value calculation process of the existing model by the model production | generation part 110 is demonstrated. The model generation unit 110 calculates an evaluation value of the existing model using the purchase survival probability “life” that is an existing model among various data included in the learning data 10. In FIG. 3, it has been described that the purchase survival probability “life” can be expressed as “life (x, t, T)”. However, the model generation unit 110 according to the first exemplary embodiment can be expressed by the following equation (1). It is assumed that the purchase survival probability “life (x, t, T, λ, μ)” is used.

  In the above formula (1), “x” indicates the number of purchases “x” shown in FIG. 3, “t” indicates the number of purchase days “t” shown in FIG. 3, and “T” The number of purchase days “T” is indicated. “Λ” and “μ” are variables. Further, the model generation unit 110 determines the likelihood function “F (x, t, T, λ, μ)” expressed by the following equation (2) in order to determine the “λ” and “μ”. Shall be used.

  In the above formulas (1) and (2), “^” indicates power and “e” indicates the base of natural logarithm. Further, the purchase survival probability “life (x, t, T, λ, μ)” in the above formula (1) can take values from “0” to “1”.

  First, the model generation unit 110 acquires the number of purchases “x”, the number of purchases “t”, and the number of first purchases “T” of the predetermined user from the learning data 10, and the acquired number of purchases “x” and the number of purchases “t” ”And the initial purchase date“ T ”are substituted into the above formula (1) to obtain“ life (λ, μ) ”where the unknowns are“ λ ”and“ μ ”.

  Subsequently, the model generation unit 110 substitutes the number of purchases “x”, the number of purchase days “t”, and the number of first purchases “T” acquired from the learning data 10 into the above equation (2), and then “λ” and “ The likelihood function F (x, t, T, λ, μ) is calculated while varying “μ” for each predetermined value set in the range of “0” to “1”. Then, the model generation unit 110 specifies “λ” and “μ” at which the likelihood function “F (x, t, T, λ, μ)” has the maximum value. Then, the model generation unit 110 sets the identified “λ” and “μ” as “λ_max” and “μ_max” respectively, and substitutes “λ_max” and “μ_max” into “life (λ, μ)”. Let “life_max”.

  The model generating unit 110 calculates “life_max” for all users whose learning data 10 includes the number of purchases “x”, the number of purchase days “t”, and the number of first purchase days “T”. Thereby, the model generation unit 110 obtains a combination of “λ_max”, “μ_max”, and “life_max” for all such users.

  Subsequently, the model generation unit 110 acquires the purchase presence / absence “real_life” from the learning data 10. Then, the model generation unit 110 calculates an evaluation value (AIC or the like) using “life_max” corresponding to each user and purchase presence / absence “real_life”. When the evaluation value is AIC, the model generation unit 110 can calculate the AIC by “−2 (log of maximum likelihood−number of free parameters)”. In this way, the model generation unit 110 calculates the evaluation value of the existing model.

  Next, (B) an alternative model generation process and an alternative model evaluation value calculation process by the model generation unit 110 will be described. First, the model generation unit 110 includes parameters “x”, “t”, “T” included in “life (x, t, T, λ, μ)” that is an existing model represented by the above formula (1). , “Λ”, and “μ”, a combination pattern (hereinafter, sometimes referred to as “alternative pattern”) of a predetermined number or more (here, “3 or more”) is determined. For example, the model generation unit 110 replaces the alternative pattern 1 = (x, t, T), the alternative pattern 2 = (x, t, λ), the alternative pattern 3 = (x, t, μ),. m = (x, t, T, λ, μ) and the like are determined.

  Subsequently, the model generation unit 110 uses the regression analysis (multiple regression analysis, logistic regression, etc.) using various data stored in the learning data 10 to display each parameter of each alternative pattern as browsing information, cart input information, etc. view_t, cart_x, etc.). Specifically, the model generation unit 110 selects the optimal browsing information (view_x, etc.) as an alternative to each parameter (x, t, T, etc.) of the alternative pattern by regression analysis, and browses each parameter selected. Find alternative formulas replaced with information.

  For example, in the case of the alternative pattern 1, since the parameters are “x”, “t”, and “T”, the model generation unit 110 represents the alternative formula “x” represented by browsing information or the like (view_t or the like). alt_x (browsing information etc.) ", an alternative formula" alt_t (browsing information etc.) "representing" t "with browsing information etc., and an alternative formula" alt_T (browsing information) representing "T" with browsing information etc. Etc.) ".

  Note that the model generation unit 110 increases the weighting of various data corresponding to purchase / non-purchase of the learning data 10 “real_life” = “1 (purchase)”, and sets the purchase / non-purchase “real_life” = “0 (no purchase)”. The regression analysis may be performed by reducing the weight of the corresponding data.

  Subsequently, the model generation unit 110 substitutes the substitute expression obtained in this way into “life (x, t, T, λ, μ)” shown in the above expression (1), thereby assigning each substitute pattern. Generate a corresponding alternative model. For example, in the case of the alternative pattern 1, with respect to “life (x, t, T, λ, μ)”, an expression that represents “x”, “t”, and “T” by browsing information etc. “Alt_t (browsing information etc.)”) is substituted, so the alternative model is “life (browsing information etc., λ, μ)”. Further, for example, in the case of the alternative pattern 2, with respect to “life (x, t, T, λ, μ)”, an expression expressing “x”, “t”, and “λ” by browsing information or the like is provided. Since it is substituted, the alternative model is “life (browsing information, T, μ)”.

  Subsequently, the model generation unit 110 acquires the browsing information or the like used in the above alternative formula from the learning data 10, and substitutes the acquired browsing information or the like into the alternative model. For example, in the case of the alternative pattern 1, the model generation unit 110 substitutes browsing information or the like into the alternative model “life (browsing information or the like, λ, μ)” so that the unknowns become “λ” and “μ”. “Life (λ, μ)” is obtained. Further, for example, in the case of the alternative pattern 2, the model generation unit 110 substitutes browsing information or the like into the alternative model alternative model “life (browsing information, T, μ)”, so that the unknowns are “T” and “ “life (T, μ)” to be “μ” is obtained.

  In addition, the model generation unit 110 substitutes the alternative formula and the browsing information acquired from the learning data 10 into the formula (2). For example, in the case of the alternative pattern 1, the model generation unit 110 substitutes the alternative formula, the browsing information, and the like into the likelihood function F (x, t, T, λ, μ) shown in the above formula (2). Then, “F (λ, μ)” where the unknowns are “λ” and “μ” is obtained. Further, for example, in the case of the alternative pattern 2, the model generation unit 110 substitutes the alternative expression, the browsing information, and the like into the likelihood function F (x, t, T, λ, μ), so that the unknown is “T "F (T, μ)" to be "and" μ ".

  Subsequently, the model generation unit 110 sets a predetermined number of unknowns of the likelihood function F (“λ” and “μ” in the case of the alternative pattern 1) in a range of “0” to “1”. The value of the likelihood function F is calculated while changing for each value. Then, the model generation unit 110 identifies an unknown having the maximum likelihood function F, and substitutes the identified unknown for an alternative model (“life (λ, μ) in the case of alternative pattern 1)”. life_max ". The model generation unit 110 calculates “life_max” for all users, and calculates an evaluation value (AIC or the like) using the calculated “life_max” and purchase presence / absence “real_life”. That is, the model generation unit 110 calculates an evaluation value (AIC or the like) of the alternative model in the same manner as the method described in (A) above.

  In this way, the model generation unit 110 generates alternative models as many as the number of alternative patterns, and calculates an evaluation value for each alternative model.

  Here, FIG. 4 shows an example of the alternative model group 30 in the first embodiment. As illustrated in FIG. 4, the model generation unit 110 generates an alternative model group 30 using various data included in the learning data 10 (see FIG. 3), and calculates an evaluation value of each alternative model. In the alternative model group 30 illustrated in FIG. 4, the model “Mm” corresponds to the alternative model corresponding to the alternative pattern m. For example, the model “M1” corresponds to an alternative model corresponding to the alternative pattern 1, and the model “M2” corresponds to an alternative model corresponding to the alternative pattern 2.

  For example, since the model “M1” has the alternative pattern 1 = (x, t, T), among the existing models, the number of purchases “x” is represented by the alternative formula “alt_x (viewing information etc.)” and the number of purchase days In this model, “t” is represented by an alternative formula “alt_t (browsing information, etc.)” and the first purchase date “T” is represented by an alternative formula “alt_T (browsing information, etc.)”. Here, the model generation unit 110 calculates “5” as the AIC indicating the evaluation value of the alternative model “M1”.

  In addition, since the model “M2” has the alternative pattern 2 = (x, t, λ), among the existing models, the number of purchases “x” is represented by the alternative formula “alt_x (viewing information etc.)” and the number of purchase days “T” is represented by an alternative expression “alt_t (browsing information etc.)” and “λ” is a model represented by an alternative expression “alt_λ (browsing information etc.)”. Here, the model generation unit 110 calculates “5.5” as the AIC indicating the evaluation value of the existing model “M2”.

  Returning to the description of FIG. 2, the accepting unit 120 accepts an available variable from the information processing apparatus 1 operated by the analyst P1. For example, the reception unit 120 receives the browsing count “view_x” illustrated in FIG. 3 as an available variable.

  The model presentation unit 130 selects an alternative model represented by the usable variable received by the receiving unit 120 and having a high evaluation value from the alternative model group 30 generated by the model generating unit 110. Then, the selected alternative model is transmitted to the information processing apparatus 1.

  Specifically, as described with reference to FIG. 4, the alternative model group 30 generated by the model generation unit 110 includes the number of purchases “x”, the number of purchase days “t”, the number of first purchase days “T”, and the survival of purchases. The explanatory variable is not only the probability “life” but also the number of views “view_x” and the like. Therefore, the model presentation unit 130 extracts, from the alternative model group 30, an alternative model that uses only the parameters included in the usable variables accepted by the accepting unit 120 as explanatory variables. Then, the model presentation unit 130 selects an alternative model having the highest evaluation value from the extracted alternative models, and transmits the selected alternative model to the information processing apparatus 1.

  Here, the process by the model presentation part 130 is demonstrated using FIG. FIG. 5 is a diagram illustrating an example of processing performed by the model presentation unit 130 according to the first embodiment. In the example illustrated in FIG. 5, the reception unit 120 receives, from the information processing apparatus 1, the number of browsing times “view_x”, the number of browsing days “view_t”, the number of first browsing days “view_T”, the browsing survival probability “view_life”, and the number of cart uses “cart_x”. ”, Cart use days“ cart_t ”, first cart use days“ cart_T ”, cart use survival probability“ cart_life ”, and purchase cycle“ buy_cycle ”are accepted as usable variables.

  In such a case, the model presentation unit 130 extracts an alternative model in which the explanatory variables are included in the usable variables received from the information processing apparatus 1 from the alternative model group 30. Here, it is assumed that the model presentation unit 130 extracts alternative models M1 to M4. The model presentation unit 130 selects an alternative model having the highest evaluation value from among the alternative models M1 to M4. When using AIC as the evaluation value as in the example shown in FIG. 5, the smaller the AIC, the higher the evaluation, and the model presentation unit 130 selects the alternative model M1 from the alternative models M1 to M4. The selected alternative model M1 and the evaluation value of the alternative model M1 are transmitted to the information processing apparatus 1. At this time, the model presentation unit 130 also transmits the evaluation value of the existing model calculated in (A) to the information processing apparatus 1.

[Processing procedure by model generator]
Next, a model generation process performed by the model generation apparatus 100 according to the first embodiment will be described with reference to FIGS. 6 and 7 are flowcharts illustrating a model generation processing procedure performed by the model generation apparatus 100 according to the first embodiment. 6 shows an evaluation value calculation process for an existing model by the model generation apparatus 100, and FIG. 7 shows an alternative model generation process and an alternative model evaluation value calculation process by the model generation apparatus 100. FIG. 8 is a flowchart illustrating an alternative model presentation processing procedure performed by the model generation apparatus 100 according to the first embodiment.

  When calculating the evaluation value of the existing model, the model generation unit 110 of the model generation device 100 uses purchase information of users 1 to n who have purchased products in the past. The model generation unit 110 first sets “n = 1” (step S101).

  Subsequently, the model generation unit 110 acquires purchase information (the number of purchases “x”, the number of purchase days “t”, the number of first purchase days “T”, etc.) from the learning data 10 (step S102). Subsequently, the model generation unit 110 substitutes the acquired purchase information into the existing model (see the above formula (1)), thereby “life (λ, μ)” where the unknowns are “λ” and “μ”. Ask for. Subsequently, the model generation unit 110 assigns the purchase information acquired from the learning data 10 to the likelihood function F (see the above equation (2)), and then the value of the likelihood function F becomes the maximum value. By substituting “λ” and “μ” into “life (λ, μ)”, the royalty of the user n is calculated (step S103).

  Subsequently, the model generation unit 110 increments “n” (step S104), and compares “n” after the increment with the number of users having purchase information (step S105). Then, when the number of users for which purchase information exists is not smaller than “n” (No at Step S105), the model generation unit 110 performs the processing procedure at Steps S102 to S104. That is, the model generation unit 110 performs the processing procedure in steps S102 to S104 for all users for which purchase information exists.

  On the other hand, if the number of users for whom purchase information exists is smaller than “n” (Yes in step S105), the model generation unit 110 determines the loyalty of each user calculated in step S103 and the correct answer corresponding to each user. An evaluation value such as AIC is calculated using the data (purchase presence / absence “real_life”) (step S106).

  As shown in FIG. 7, the model generation unit 110 of the model generation apparatus 100 determines an alternative pattern that is a combination of a predetermined number or more parameters from the parameters included in the existing model (step S201). Here, it is assumed that the model generation unit 110 determines alternative patterns 1 to m (a total of m types of alternative patterns). In such a case, the model generation unit 110 first sets “m = 1” (step S202).

  Subsequently, the model generation unit 110 obtains an alternative expression in which each parameter of the alternative pattern m is represented by browsing information or the like (step S203). At this time, the model generation unit 110 selects optimal browsing information or the like as a substitute for each parameter by regression analysis using various data stored in the learning data 10, and replaces each parameter with the selected browsing information or the like. To find an alternative formula. Subsequently, the model generation unit 110 generates an alternative model corresponding to the alternative pattern m by substituting the alternative formula into the existing model (see the above formula (1)) (step S204).

  Subsequently, the model generation unit 110 calculates loyalty using various information of the users 1 to n whose various information exists in the learning data 10. Specifically, the model generation unit 110 first sets “n = 1” (step S205), and acquires browsing information and the like of the user n from the learning data 10 (step S206). At this time, the model generation unit 110 acquires browsing information or the like that has become a variable in the alternative formula obtained in step S203. Subsequently, the model generation unit 110 substitutes the acquired browsing information or the like into the alternative model and likelihood function F generated in step S204. Since there are similar unknowns in the alternative model and the likelihood function F, the model generation unit 110 substitutes the unknown when the value of the likelihood function F reaches the maximum value into the alternative model, and the royalties of the user n The tee is calculated (step S207).

  Subsequently, the model generation unit 110 increments “n” (step S208), and compares “n” after the increment with the number of all users who have various information in the learning data 10 (step S209). Then, when the total number of users is not smaller than “n” (No at Step S209), the model generation unit 110 performs the processing procedure at Steps S205 to S208. That is, the model generation unit 110 performs the processing procedure in steps S205 to S208 for all users.

  On the other hand, when the total number of users is smaller than “n” (Yes at Step S209), the model generation unit 110 determines the loyalty of each user calculated at Step S207 and the correct data corresponding to each user (purchase presence / absence “ real_life ") is used to calculate an evaluation value such as AIC (step S210).

  Subsequently, the model generation unit 110 increments “m” (step S211), and determines whether the m types of alternative patterns determined in step S201 have been processed (step S212). And the model production | generation part 110 performs the process sequence in said step S203-S211, when not processing about all the alternative patterns (step S212 negative). On the other hand, if the model generation unit 110 has processed all the alternative patterns (Yes at step S212), the model generation unit 110 ends the processing.

  Further, as illustrated in FIG. 8, the reception unit 120 of the model generation device 100 monitors whether an available variable has been received (step S301). Here, when the accepting unit 120 does not accept the usable variable (No at Step S301), the accepting unit 120 waits until the usable variable is accepted.

  On the other hand, when the usable variable is accepted by the accepting unit 120 (Yes in step S301), the model presenting unit 130 converts the parameter included in the usable variable from the alternative model generated by the model generating unit 110 to the explanatory variable. The alternative model is extracted (step S302). Subsequently, the model presentation unit 130 selects an alternative model having the highest evaluation value from the extracted alternative models (step S303).

  Then, the model presentation unit 130 transmits the selected alternative model, the evaluation value of the alternative model calculated in step S210, and the evaluation value of the existing model calculated in step S106 to the information processing apparatus 1 ( Step S304).

[Effect of Example 1]
As described above, the model generation apparatus 100 according to the first embodiment uses an explanatory variable other than the purchase history for an existing model that obtains an objective variable that is a user's loyalty using the purchase history of the product by the user as an explanatory variable. A plurality of alternative models that are models replaced with other history information are generated. Then, the model generation apparatus 100 calculates an evaluation value indicating the suitability of the other history information with respect to the alternative model for each generated alternative model. Then, the model generation apparatus 100 receives input of usable variables that are usable explanatory variables, extracts a replacement model in which the other history information is included in the usable variable from a plurality of replacement models, and extracts the replacement The alternative model with the highest evaluation value is presented.

  As a result, the model generation device 100 according to the first embodiment can present to the analyst P1 an alternative model having history information (browsing information or the like) other than the purchase history as an explanatory variable. As a result, the analyst P1 can measure loyalty even for a user who has not purchased a product in the past.

  Further, the model generation apparatus 100 according to the first embodiment calculates an evaluation value of the existing model, and also presents the calculated evaluation value of the existing model.

  Thereby, the model generation apparatus 100 according to the first embodiment can provide the analyst P1 with the evaluation value of the existing model to be compared with the evaluation value of the alternative model. As a result, the analyst P1 can grasp the degree of accuracy with which the alternative model can measure the royalty as compared with the existing model.

  Incidentally, the model generation method and the model generation apparatus according to the present invention may be implemented in various different forms other than the above-described embodiments. Accordingly, in the second embodiment, another embodiment of the model generation method and the model generation apparatus according to the present invention will be described.

[Evaluation value of existing model]
In the first embodiment, the example in which the model generation apparatus 100 presents the alternative model, the evaluation value of the alternative model, and the evaluation value of the existing model is shown. However, the model generation apparatus 100 may present only the alternative model, or may present only the alternative model and the evaluation value of the alternative model.

[Processing object]
Moreover, in the said Example 1, when the model production | generation apparatus 100 calculated the evaluation value of the existing model, the example which uses the purchase information of all the users who have purchased goods in the past was shown. However, the model generation apparatus 100 may calculate the evaluation value of the existing model using purchase information for a predetermined number of users without using purchase information for all users.

  Similarly, in the first embodiment, the example in which the model generation apparatus 100 uses the browsing information of all users when the evaluation value of the alternative model is calculated is shown. However, the model generation apparatus 100 may calculate the evaluation value of the alternative model using the browsing information of a predetermined number of users without using the browsing information of all users.

[Alternative pattern]
In the first embodiment, the example in which the model generation apparatus 100 determines an alternative pattern that is a combination pattern of parameters has been described. At this time, the model generation apparatus 100 may use all combinations of a predetermined number or more of the parameters included in the existing model as alternative patterns, or parameters with high correlation (for example, view_life, view_life, t, T Etc.) may be used as an alternative pattern only.

[Measurement of loyalty]
In the first embodiment, the example in which the model generation device 100 transmits the alternative model to the information processing device 1 has been described. However, when the model generation apparatus 100 accepts various types of information (usable variables) of the user whose royalty is to be measured, the model generation apparatus 100 measures the royalty using the alternative model and sends the measurement result to the information processing apparatus 1. You may send it.

[Embodiment]
Moreover, in the said Example 1, the example which the model production | generation apparatus 100 produces | generates the alternative model for measuring the loyalty of the user who uses a shopping site was shown. However, the system to which the model generation device 100 is applied is not limited to a shopping site. For example, the model generation apparatus 100 can also be applied when generating an alternative model for measuring the loyalty of a customer such as a retail store.

[System configuration]
In addition, among the processes described in the above embodiment, all or a part of the processes described as being automatically performed can be manually performed, or the processes described as being performed manually can be performed. All or a part can be automatically performed by a known method. In addition, the processing procedures, specific names, and information including various data and parameters shown in the document and drawings can be arbitrarily changed unless otherwise specified. For example, the various types of information illustrated in FIGS. 3 to 5 are merely examples, and can be changed to arbitrary information.

  Further, each component of each illustrated apparatus is functionally conceptual, and does not necessarily need to be physically configured as illustrated. In other words, the specific form of distribution / integration of each device is not limited to that shown in the figure, and all or a part thereof may be functionally or physically distributed or arbitrarily distributed in arbitrary units according to various loads or usage conditions. Can be integrated and configured. For example, in the example illustrated in FIGS. 1 and 2, the model generation apparatus 100 has the learning data 10. However, the learning data 10 may be held by a device other than the model generation apparatus 100. In such a case, the model generation device 100 acquires the learning data 10 from another device.

[program]
It is also possible to create a program in which the processing executed by the model generation apparatus 100 described in the first and second embodiments is described in a language that can be executed by a computer. For example, it is possible to create a model generation program in which processing executed by the model generation apparatus 100 according to the first embodiment is described in a language that can be executed by a computer. In this case, when the computer executes the model generation program, the same effects as in the first embodiment can be obtained. Further, the model generation program is recorded on a computer-readable recording medium, and the model generation program recorded on the recording medium is read by the computer and executed, thereby realizing the same processing as in the first embodiment. Also good. Hereinafter, an example of a computer that executes a model generation program that realizes the same function as the model generation apparatus 100 illustrated in FIG. 2 will be described.

  FIG. 9 is a diagram illustrating a computer 1000 that executes a model generation program. As illustrated in FIG. 9, the computer 1000 includes, for example, a memory 1010, a CPU 1020, a hard disk drive interface 1030, a disk drive interface 1040, a serial port interface 1050, a video adapter 1060, and a network interface 1070. These units are connected by a bus 1080.

  The memory 1010 includes a ROM (Read Only Memory) 1011 and a RAM (Random Access Memory) 1012 as illustrated in FIG. The ROM 1011 stores a boot program such as BIOS (Basic Input Output System). The hard disk drive interface 1030 is connected to the hard disk drive 1031 as illustrated in FIG. The disk drive interface 1040 is connected to the disk drive 1041 as illustrated in FIG. For example, a removable storage medium such as a magnetic disk or an optical disk is inserted into the disk drive. The serial port interface 1050 is connected to a mouse 1051 and a keyboard 1052, for example, as illustrated in FIG. The video adapter 1060 is connected to a display 1061, for example, as illustrated in FIG.

  Here, as illustrated in FIG. 9, the hard disk drive 1031 stores, for example, an OS 1091, an application program 1092, a program module 1093, and program data 1094. That is, the model generation program is stored in, for example, the hard disk drive 1031 as a program module in which a command to be executed by the computer 1000 is described. For example, a model generation procedure for executing information processing similar to that of the model generation unit 110 illustrated in FIG. 2, a reception procedure for executing information processing similar to that of the reception unit 120, and information processing similar to that of the model presentation unit 130 are executed. A program module 1093 in which a model presentation procedure is described is stored in the hard disk drive 1031.

  Various data held by the model generation apparatus 100 described in the above embodiment is stored as program data in, for example, the memory 1010 or the hard disk drive 1031. Then, the CPU 1020 reads the program module 1093 and the program data 1094 stored in the memory 1010 and the hard disk drive 1031 to the RAM 1012 as necessary, and executes a model generation procedure, a reception procedure, and a model presentation procedure.

  Note that the program module 1093 and the program data 1094 related to the model generation program are not limited to being stored in the hard disk drive 1031, but are stored in, for example, a removable storage medium and read out by the CPU 1020 via the disk drive or the like. Also good. Alternatively, the program module 1093 and the program data 1094 related to the model generation program are stored in another computer connected via a network (LAN (Local Area Network), WAN (Wide Area Network), etc.), and the network interface 1070 is stored. Via the CPU 1020.

DESCRIPTION OF SYMBOLS 100 Model production | generation apparatus 110 Model production | generation part 120 Reception part 130 Model presentation part

Claims (4)

A model generation method executed by a model generation device,
An alternative model in which the explanatory variable is replaced with other history information other than the purchase history with respect to an existing model which is a model for obtaining an objective variable which is the user's loyalty using the purchase history of the product by the user as an explanatory variable A model generation process for generating multiple models;
For each alternative model generated by the model generation step, an evaluation value calculation step for calculating an evaluation value indicating the degree of fitness of the other history information with respect to the alternative model;
A reception process that accepts input of usable variables that are usable explanatory variables;
An alternative model in which the other history information is included in the usable variable accepted by the acceptance step is extracted from a plurality of alternative models generated by the model generation step, and the evaluation value calculation step is performed among the extracted alternative models And a model presentation step of presenting an alternative model having the highest calculated evaluation value. The evaluation value calculation step includes:
Calculating an evaluation value indicating the degree of conformity of the purchase history with respect to the existing model;
The model presentation step includes
The model generation method according to claim 1, further presenting an evaluation value of the existing model calculated by the evaluation value calculation step. An alternative model in which the explanatory variable is replaced with other history information other than the purchase history with respect to an existing model which is a model for obtaining an objective variable which is the user's loyalty using the purchase history of the product by the user as an explanatory variable. A model generator that generates multiple models;
For each alternative model generated by the model generation unit, an evaluation value calculation unit that calculates an evaluation value indicating the degree of fitness of the other history information with respect to the alternative model;
A reception unit that accepts input of usable variables that are usable explanatory variables;
An alternative model in which the other history information is included in the usable variable received by the receiving unit is extracted from a plurality of alternative models generated by the model generating unit, and the evaluation value calculating unit out of the extracted alternative models A model generation apparatus comprising: a model presentation unit that presents an alternative model having the highest calculated evaluation value. The evaluation value calculation unit
Calculating an evaluation value indicating the degree of conformity of the purchase history with respect to the existing model;
The model presentation unit
The model generation apparatus according to claim 3, further presenting an evaluation value of the existing model calculated by the evaluation value calculation unit.

Images