JP6101620B2 - Purchase forecasting apparatus, method, and program - Google Patents

Description

  The present invention relates to a purchase prediction apparatus, method, and program, and more particularly, to a purchase prediction apparatus, method, and program for predicting a customer's next purchase based on past purchase behavior.

  The purchase forecasting technology for products that the customer is interested in is an important elemental technology when considering the purchase of products and advertising strategies.

  As conventional technology in this field, there is one that pays attention to product prices and customer purchasing behavior. For example, Non-Patent Document 1 models how product discounts affect short-term and long-term sales. However, this method predicts overall sales, and there is a problem that it is not known who the customer purchases the product at what price.

  On the other hand, Non-Patent Document 2 discloses a method of modeling customer purchasing behavior with purchased products and their prices. However, this method is a model that assumes a price range that is easy to purchase only for products, and there is a problem that it is not possible to explicitly model which price range a customer is likely to purchase.

  Non-Patent Document 3 discloses a technique for estimating a potential topic structure hidden behind a document set by learning based on variational Bayes.

Nijs VR, Dekimpe MG, Steenkamps JBE, Hanssens DM, “The category-demand effects of price promotions,” Marketing Science 20 (1), pp.1-22 (2001). Tomoharu Iwata and Hiroshi Sawada, “Topic model for analyzing purchase data with price information,” Data Mining and Knowledge Discovery, 26, 3, pp.559-573 (May 2013). David M. Blei, Andrew Y. Ng, Michael I. Jordan, “Latent Dirichlet Allocation” The Journal of Machine Learning Research, 3, pp.993-1022 (March 2003).

  As described above, in the prior art, the customers are collected as a customer group, not as individual units, and an analysis related to the purchase behavior of the customer group is performed. For this reason, it has not been possible to analyze, for example, which price range will be purchased next for each customer.

  The present invention has been made in view of the above circumstances, and an object thereof is to provide a purchase prediction apparatus, method, and program capable of predicting purchase behavior for each customer with higher accuracy.

  In order to achieve the above object, a purchase prediction apparatus according to the present invention includes customer identification information for identifying a customer, product identification information for identifying a product purchased by the customer, and the customer purchased An acquisition unit that acquires a set of customer purchase information including a product price of the product, and the product purchased by the customer for each of the customers based on the set of customer purchase information acquired by the acquisition unit Product price information representing each product price, purchase price information representing each product price when the product is purchased for each of the products, and the customer purchased for each of the customers Data generation means for generating purchase quantity information representing each product identification information of the product, and for each of the customers, the product price of the customer generated by the data generation means Based on the information, an internal reference price learning means for estimating a probability distribution of the product price as an internal reference price that is a price range that the customer considers purchasing, and for each of the products, the data generation means Based on the generated purchase price information of the product, the product price range learning means for estimating a probability distribution of the product price as the price range of the product, and the data for each combination of the customer and the product A probability distribution representing the probability of belonging to each topic based on the customer's interest, obtained based on the number-of-purchase information generated by the generating means, and a probability of purchasing the product in the topic for each topic A customer interest learning means for estimating a probability that the customer of the combination purchases the product based on a probability distribution; and the product For each, a similarity for calculating the similarity between the probability distribution of the product price estimated as the internal reference price for the customer to be predicted and the probability distribution of the product price estimated as the purchase price information for the product A degree calculation means, for each of the products, the similarity to the product calculated by the similarity calculation means, and the probability of purchasing the product estimated for the combination of the product and the customer to be predicted; And a score calculation means for calculating a score indicating the possibility that the prediction target customer purchases the product.

  A purchase prediction method according to the present invention includes an acquisition unit, a data generation unit, an internal reference price learning unit, a product price range learning unit, a customer interest learning unit, a similarity calculation unit, and a score calculation unit. A purchase prediction method in a purchase prediction apparatus including: customer identification information for identifying a customer; product identification information for identifying a product purchased by the customer; and the customer purchased by the acquisition unit A set of customer purchase information including a product price of a product is acquired, and the customer purchases each of the customers based on the set of customer purchase information acquired by the acquisition unit by the data generation unit. Product price information representing the product price of each of the products, purchase price information representing the product price of each of the products when the product is purchased, and the customer For each of the products, the purchase number information representing the product identification information of each of the products purchased by the customer is generated, and the data is generated by the data generation unit for each of the customers by the internal reference price learning unit. Based on the product price information of the customer, the probability distribution of the product price is estimated as an internal reference price that is a price range that the customer considers purchasing, and each of the products is estimated by the product price range learning means. The probability distribution of the product price is estimated as the price range of the product based on the purchase price information of the product generated by the data generation unit, and the customer and the product are estimated by the customer interest learning unit. For each of the combinations, based on the customer's interest determined based on the purchase quantity information generated by the data generation means. The probability that the customer of the combination purchases the product is estimated based on a probability distribution representing the probability belonging to each topic and a probability distribution representing the probability that the product in the topic is purchased for each topic. , For each of the products, the product price probability distribution estimated as the internal reference price for the prediction target customer and the product price estimated as the purchase price information for the product. The degree of similarity with the probability distribution of the product is calculated, and the score calculation unit calculates the similarity for the product calculated by the similarity calculation unit for each of the products, and the product and the prediction target customer. Based on the probability of purchasing the product estimated for the combination, the prediction target customer purchases the product. Calculate a score indicating the likelihood

  The program according to the present invention is a program for causing a computer to function as each unit of the purchase prediction apparatus.

  As described above, according to the purchase prediction apparatus, method, and program of the present invention, based on a set of customer purchase information including a combination of a product purchased by a customer and its price, Estimate the probability distribution of the price range, the probability distribution of the price for each purchased product, and the probability distribution of the interest of the customer's product, and based on these estimated probability distributions, By calculating a score representing the possibility of purchasing the product, it is possible to predict the purchase behavior for each customer with higher accuracy.

It is a block diagram which shows the functional structure of the purchase prediction apparatus of this Embodiment. It is a figure which shows the example of an output of the purchase action prediction model learning part in this Embodiment. It is a flowchart which shows the content of the purchase action prediction model learning routine in the purchase prediction apparatus which concerns on embodiment of this invention. It is a flowchart which shows the content of the purchase action prediction routine in the purchase prediction apparatus which concerns on embodiment of this invention. It is a figure which shows an example of the data structure of various data. It is a figure which shows the example of a purchase prediction output when performing the purchase prediction process routine in the purchase prediction apparatus which concerns on embodiment of this invention.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

  The purchase prediction apparatus 100 according to the present embodiment includes a CPU (Central Processing Unit), a RAM (Random Access Memory), and a ROM (Read Only Memory) that stores a program for executing a purchase prediction processing routine to be described later. It is composed of a computer with As shown in FIG. 1, this computer can be functionally represented by a configuration including an acquisition unit 10, a calculation unit 20, a storage unit 30, and an output unit 40.

  The acquisition unit 10 acquires a set of customer purchase information indicating a combination of a product purchased by a customer, its price, and information of a customer who purchased the product.

  The calculation unit 20 includes a purchase behavior prediction model learning unit 21 and a purchase behavior prediction unit 22. The purchase behavior prediction model learning unit 21 includes a data generation unit 21a, an internal reference price learning processing unit 21b, a product price range learning processing unit 21c, and a customer interest learning processing unit 21d. A similarity calculation unit 22a and a score calculation unit 22b are provided.

<Purchasing Behavior Prediction Model Learning Department>

  In the purchase behavior prediction model learning unit 21, a set of customer purchase information acquired by the acquisition unit 10 is input, and as a result of learning, a probability distribution of a price range (hereinafter referred to as an internal reference price) of a product purchased for each customer. The probability distribution of the price for each purchased product and the probability distribution of the interest of the customer's product are output.

First, the data generation unit 21a in FIG. 1 sets each set L _i = <consisting of the customer id (variable u _i ), the product id (variable b _i ), and the product price (variable v _i ) acquired by the acquisition unit 10. A set of customer purchase information AllLogs = <L ₁ , L ₂ ,..., L _i ,..., L _M > configured by u _i , b _i , v _i > Here, M represents the total number of purchases where K types of products k (k = 1,..., K) are purchased by all N customers n (n = 1,..., N). . Since the same customer may purchase the same product many times, M is not limited to the value of (N × K) and may take a value of (N × K) or more.

Then, in the data generation unit 21a, product price information UserLogs = <U ₁ ,..., U _N >, U _n = {v _i | u _i = n, i∈1,..., M}, purchase price information indicating the purchased price for each product k ItemLogs = <B ₁ , ..., B _K >, B _k = {v _i | b _i = k, _i ∈ 1,..., M}, and purchase quantity information UserPurchase = <Upurchase ₁ ,..., Upurchase _N >, Upurchase _n = {b _i | u _i = n, i∈1,..., M} is generated.

  Then, the product price information UserLogs representing the product price purchased for each customer n is output to the internal reference price learning processing unit 21b, and the purchase price information ItemLogs representing the price purchased for each product k is output to the product price range learning processing unit. It outputs to 21c, and the purchase number information UserPurchase of the purchased goods for every customer n is output to the customer interest learning process part 21d.

The internal reference price learning processing unit 21b accepts product price information UserLogs = <U ₁ ,..., U _N > for each customer n output from the data generation unit 21a. Then, about the product price information U _n that each customer n has purchased, to estimate the internal reference price, as the probability distribution Pinner _n take the price Price in variable. For example, a mixed Gaussian distribution is used for estimating the probability distribution Pinner _n . As a result, the probability distribution Pinner _n of the internal reference price of customer n is, as a weighted sum of the one-dimensional normal distribution of a finite number is the probability distribution to take the price Price in variable (I number), (1) as of the formula It is expressed in

Here, N (•) represents a normal distribution. Further, μ _i and σ _i ² are the average value and variance value of the normal distribution i for representing the price range of the internal reference price of the customer n, respectively, and π _i is a multiplier of the normal distribution i.

The distribution used for estimating the probability distribution Pinner _n is not limited to the mixed Gaussian distribution, and other probability distributions such as a lognormal distribution and a Weibull distribution may be used.

In this way, the internal reference price learning processing unit 21b is a model curve representing the probability that the customer n purchases an arbitrary product price for each customer n, and the probability distribution Pinner of the internal reference price of the customer n. _n is estimated and output to the purchase behavior prediction unit 22.

Next, the product price range learning processing unit 21c receives purchase price information ItemLogs = <B ₁ ,..., B _K > representing the purchased price for each product k output from the data generation unit 21a.

After that, for the purchased price information B _k for each product k, the price range of the product is estimated as a probability distribution Pb _k taking the price Price as a variable. For example, a mixed Gaussian distribution is used for estimating the probability distribution Pb _k . Table as a result, the probability distribution Pb _k of the price range of commodity k is, as a weighted sum of the one-dimensional normal distribution of a finite number (J pieces) is the probability distribution to take the price Price in variable, equation (2) Is done.

Here, γ _j and δ _j ² are the average value and variance value of the normal distribution j for representing the price range of the product k, respectively, and ζ _j is a multiplier of the normal distribution j.

The distribution used for estimating the probability distribution Pb _k is not limited to the mixed Gaussian distribution, and other probability distributions such as a lognormal distribution and a Weibull distribution may be used.

In this way, the product price range learning processing unit 21c estimates the probability distribution Pb _k of the price range of the product k, which is a model curve representing the probability that the product k is purchased at a certain price for each product k. Output to the behavior prediction unit 22.

The customer interest learning processing unit 21d receives the number-of-purchase information UserPurchase = <Upurchase ₁ ,..., Upurchase _N > of products purchased for each customer n output from the data generation unit 21a. Thereafter, for each combination of the customer n and the product k, the probability Q that the customer n purchases the product k is estimated, and the probability distribution that the customer based on the customer's interest purchases the product is estimated.

  LDA (Latent Dirichlet Allocation) described in Non-Patent Document 3 is used for customer interest in products. With LDA, a probability distribution θ (z | n) of each product set (topic) z based on the interest of customer n and a probability distribution φ (k | z) of product k in each topic z are obtained. Here, z is the customer's interest in the behavior of customer n purchasing product k. In other words, by using LDA, the product of the degree of interest z of each customer n θ (z | n) and the ease of purchase of the product k in each interest z φ (k | z) The purchase of the item k can be estimated.

  The probability Q (k | n) that the customer n based on the customer's interest purchases the product k is expressed by the following equation (3).

  As described above, the customer interest learning processing unit 21d estimates, for each combination of the customer n and the product k, the probability Q (k | n) that the customer n purchases the product k based on the customer's interest. The probability distribution that the customer based on the purchase of the product is output to the purchase behavior prediction unit 22.

FIG. 2 is a diagram illustrating an output example of the purchase behavior prediction model learning unit 21. 2 (A) is estimated in the internal reference price learning processing section 21b, illustrates an example of a probability distribution Pinner _{user_1} the internal reference price customer user_1, FIG. 2 (B), commodity price learning processing unit FIG. _2C shows the probability distribution Pb B_1 of the price range of the product B_1 estimated in 21c, and FIG. _2C shows that the customer user_1 based on the customer's interest estimated by the customer interest learning processing unit 21d purchases each product k. The probability distribution Q (k | user_1) is shown.

As can be seen from FIGS. 2A to 2C, the probability distribution _{Piner user_1} and the probability distribution Pb _{B_1} are continuous probability distributions, and the probability Q (k | user_1) follows a discrete probability distribution.

<Purchase Behavior Prediction Department>

  The purchase behavior prediction unit 22 calculates a score representing the possibility that the customer corresponding to the customer id will purchase each product, using the customer id to be predicted as an input.

First, the similarity calculation unit 22a obtains a probability distribution Pinin _nin of the customer's internal reference price corresponding to the customer id: n _in based on the input customer id: n _in . Next, for all product id: k (kε1,..., K), the probability distribution Pb _k of the price range of the product _k and the probability distribution Pinin _nin of the internal reference price of the customer n _in Calculate the similarity of.

  The calculation of the similarity is expressed as in equation (4) using KL Divergence.

Thus, since Pinner _nin and Pb _k are probability distributions with respect to the price Price, KL Divergence is expressed as an integral function using the price Price as shown in the equation (4). Note that the calculation of the similarity KL (Pinner _nin , Pb _k ) between the probability distribution Pb _k and the probability distribution Pinin _nin is not limited to KL Divergence, and for example, Cauchy Schwarz Divergence, Jensen Shannon Divergence, or the like may be used.

Then, the customer purchase forecast value sim (n _in , k) is expressed as in equation (5) using KL Divergence.

As described above, the similarity calculation unit 22a predicts the customer purchase prediction that the purchase behavior of the customer is attributed to the price of the product for all the product id: k (kε1,..., K). The value sim (n _in , k) is calculated and output to the score calculation unit 22b.

The score calculation unit 22b receives the customer purchase prediction value sim (n _in , k) for each product k output from the similarity calculation unit 22a. Then, for each product k, the probability Q (k) that the customer n _in based on the customer's interest obtained by the customer interest learning processing unit 21d purchases the product k in the customer purchase forecast value sim (n _in , k). In consideration of | n _in ), score Score (k | n _in ) representing the possibility that customer n _in purchases product k is calculated according to equation (6). The process of adding the probability Q (k | n _in ) to the customer purchase forecast value sim (n _in , k) is called smoothing.

Here, α is a process for smoothing and takes a value of 0 ≦ α ≦ 1. By adjusting the value of α, the proportion of purchase behavior based on the price of the product _in the score Score (k | n _in ) and the proportion of purchase behavior based on interest in the product are adjusted.

Here, the score Score (k | n _in ) is smoothed by the linear interpolation expressed by the equation (6), but a smoothing method such as nonlinear interpolation or Unigram Rescaling may be used.

Then, the output unit 40 outputs a score Score (k | n _in ) that is predicted for each product k by the purchase behavior prediction unit 22 and represents the possibility that the customer n _in purchases the product k.

  Next, the operation of purchase prediction apparatus 100 according to the present embodiment will be described.

  FIG. 3 is a flowchart of a purchase behavior prediction model learning routine executed in the purchase prediction apparatus 100 according to the present embodiment.

First, in step S101, customer id, item id, and customer purchase information set AllLogs = _{<L 1, L 2} of that with each set _{L i} element consisting of the price of goods, ..., _{L i,} ... , L _M > and store it in the storage unit 30. The customer purchase information 50 shown in FIG. 5 shows the data structure of AllLogs.

Next, in step S102, product price information UserLogs = <U ₁ ,..., U _N > representing the product price purchased for each customer n is generated from the set AllLogs of customer purchase information stored in the storage unit 30. And stored in the storage unit 30.

Next, in step S103, purchase price information ItemLogs = <B ₁ ,..., B _K > representing the purchase price for each product is generated from the set AllLogs of customer purchase information stored in the storage unit 30. And stored in the storage unit 30.

Next, in step S104, the purchase quantity information UserPurchase = <Upurchase ₁ ,..., Upurchase _N > of the products purchased for each customer is generated from the set AllLogs of customer purchase information stored in the storage unit 30 and stored. Store in unit 30.

  In the purchase behavior prediction model learning routine of the purchase prediction apparatus 100 according to the present embodiment, the product price information UserLogs for each customer, the purchase price information ItemLogs for each product, and the purchase number information UserPurchase for the product purchased for each customer. Although generated, there is no limitation on the generation order of these pieces of information. For example, the pieces of information may be generated in the order of UserPurchase, ItemLogs, and UserLogs.

In step S105, the product price information UserLogs for each customer generated in step S102 is read from the storage unit 30, and for each customer n, based on the product price information of the customer n, the customer n shown in the above equation (1). The probability distribution Pinner _n of the internal reference price is estimated and stored in the storage unit 30. The customer's internal reference price 52 shown in FIG. 5 is an example of the probability distribution Pinner _n calculated in step S105.

Here, the arrangement of the parameter 1, parameter 2,..., Parameter r of the customer's internal reference price 52 defines, for example, each of a plurality of one-dimensional normal distributions (multiplier π _i , average value μ _i , A set of dispersion values σ _i ² ) is repeatedly shown.

Next, in step S106, the purchase price information ItemLogs for each product generated in step S103 is read from the storage unit 30, and for each product k, based on the purchase price information of the product k, the product k shown in the above equation (2). The probability distribution Pb _k of the price range is estimated and stored in the storage unit 30. The price range 54 of the product shown in FIG. 5 is an example of the probability distribution Pb _k calculated in step S106.

Here, the arrangement of the parameter 1, the parameter 2,..., The parameter s of the product price range 54 defines, for example, each of a plurality of one-dimensional normal distributions as in the data structure of the internal reference price 52. A set of (multiplier ζ _j , average value γ _j , variance value δ _j ² ) is repeatedly shown.

  Next, in step S107, the purchase quantity information UserPurchase of the product purchased for each customer generated in step S104 is read from the storage unit 30, and each interest of the customer n is obtained by the learning method based on LDA shown in Non-Patent Document 3. A probability distribution θ (z | n) representing the degree of z and a probability distribution φ (k | z) representing the ease of purchase of the product k for each interest z are calculated.

For example, the process of assigning a topic (interest) z to each purchase data of each customer n indicated by the element Upurchase _N of the purchase quantity information UserPurchase is repeatedly performed until the likelihood converges, and finally obtained. The probability distribution θ (z | n) and the probability distribution φ (k | z) may be calculated based on the topic assignment to each of the purchase data of each customer n.

  Then, for each combination of customer n and product k, according to the above equation (3), based on probability distribution θ (z | n) for customer n and probability distribution φ (k | z) for product k The probability Q (k | n) that the customer n based on the customer's interest calculates the product k is calculated, and the probability Q (k | n) for each combination is calculated as the probability that the customer based on the customer's interest purchases the product. The distribution is stored in the storage unit 30. The product purchase probability 56 based on the customer's interest shown in FIG. 5 is an example of the probability Q (k | n) calculated in step S107.

In the purchase prediction processing routine of the purchase prediction apparatus 100 according to the present embodiment, the probability distribution Pinn _n , the probability distribution Pb _k , and the probability Q (k | n) are generated in this order. The order is not limited, and for example, the distribution may be generated in the order of probability Q (k | n) distribution, probability distribution Pb _k , and probability distribution Pinn _n .

By the above process, from the set AllLogs of customer purchase information, customer probability distribution of the internal reference price of each _n Pinner _n, commodity probability distribution Pb k of the price range of each _k, and customers n commodity k based on the interest of the customer The distribution of the probability Q (k | n) of purchasing is calculated.

  On the other hand, FIG. 4 is a flowchart of a purchase behavior prediction routine executed in the purchase prediction apparatus 100 according to the present embodiment. This flowchart is performed after the purchase behavior prediction model learning routine shown in FIG.

First, in step S <b> 201, customer id: n _in which is a target of purchase prediction for a product is acquired and stored in the storage unit 30. In this case, the customer id: n _in to be acquired may be one or plural.

Next, in step S202, the probability distribution Pinin _nin of the customer's internal reference price corresponding to the customer id: n _in acquired in step S201 is read from the storage unit 30, and all the products calculated in step S106 of FIG. The probability distribution Pb k of the price range of _k is read from the storage unit 30.

Then, for each product k, the probability distribution Pb according to the above equation (4) based on the probability distribution Pinin _nin corresponding to the customer id: n _in acquired in step S201 and the probability distribution Pb _{k of the} product k. _The degree of similarity KL (Pinner _nin , Pb _k ) between _k and the probability distribution Pinin _nin of the customer's internal reference price is calculated.

Further, for each product k, based on the calculated similarity KL (Pinner _nin , Pb _k ), the customer purchase forecast value sim (n _in , k) for the product k is calculated according to the above equation (5) and stored. Store in unit 30.

In step S203, for each product k, the real number α (0 ≦ α ≦ 1) stored in the storage unit 30, sim (n _in , k) for the product k calculated in step S202, step S107 in FIG. Based on Q (k | n _in ) corresponding to the product k and the customer id: n _in calculated in step (2), the customer corresponding to the customer id: n _in may purchase the product k according to the above equation (6). The score Score (k | n _in ) representing is calculated and stored in the storage unit 30. For example, the score Score (k | n _in ) for each product k stored in the storage unit 30 is displayed on a display unit (not shown) provided in the purchase prediction device 100. The product purchase score 58 shown in FIG. 5 is an example of the score calculated in step S202.

The output destination of the score Score (k | n _in ) is not limited to the display unit. For example, the score Score (k | n _in ) is output to an information terminal (not shown) via a communication line (not shown), and the information (not shown) score terminal score (k | n _in) may be displayed, also the score in an image forming apparatus such as a printer score (k | n _in) may be output.

<Results>

  FIG. 6 illustrates a case where the purchase prediction processing routine shown in FIG. 4 is executed after the purchase behavior prediction model learning routine shown in FIG. 3 of the purchase prediction apparatus 100 according to the present embodiment is executed. It is the figure which showed the example of an output of score Score (k | user_1) at the time of acquiring customer id: user_1 in the process of step S201. In this case, as shown in FIG. 6, the score for each product k is output together with the product id to the customer id: user_1.

  Here, the score is a non-negative value, and the larger the score value, the higher the possibility that the customer user_1 purchases the product indicated by the corresponding product id at the next purchase opportunity. Further, by rearranging the score values in descending order, that is, in order from the largest score value to the smallest value, it is also possible to output products that are likely to be purchased by the customer user_1 in a ranking format. Needless to say, the score output order is not limited to descending order, and the score values may be rearranged in the order in accordance with the purpose.

  As described above, the purchase prediction apparatus 100 can analyze which product is to be purchased next for each customer, and can predict purchase behavior for each customer with higher accuracy.

  As described above, according to the purchase prediction apparatus according to the embodiment of the present invention, the purchase behavior of the customer is described by the combination of the product purchased by the customer and its price, and the set is received as an input. Then, based on the received information, estimate the probability distribution of the internal reference price of the product purchased for each customer, the probability distribution of the price for each purchased product, and the probability that the customer will purchase the product based on the customer's interest To do. Then, using the internal reference price for each customer and the probability distribution of the price for each product as input, the customer purchase forecast value based on the price of the product is calculated, and the customer purchases the product based on the customer forecast value and the customer's interest The purchase behavior for each customer can be predicted with higher accuracy by smoothing the probability of performing and outputting a score representing the possibility that the customer will purchase next for each product.

  This makes it possible to optimize product purchases and advertising strategies for each customer, as well as what price range each customer is likely to purchase, and at what price range each product is likely to be purchased. Such information can be grasped, and the purchase behavior of the customer and the brand value for each product can be grasped.

  Note that the present invention is not limited to the above-described embodiment, and various modifications and applications are possible without departing from the gist of the present invention.

  For example, the purchase forecasting apparatus described above has a computer system inside, but the “computer system” includes a homepage providing environment (or display environment) if a WWW system is used.

  In the present specification, the embodiment has been described in which the program is installed in advance. However, the program can be provided by being stored in a computer-readable recording medium.

DESCRIPTION OF SYMBOLS 10 Input part 20 Calculation part 21 Purchasing behavior prediction model learning part 22 Purchasing behavior prediction part 30 Storage part 40 Output part 100 Purchasing prediction apparatus

Claims (3)

Acquisition means for acquiring a set of customer purchase information including customer identification information for identifying a customer, product identification information for identifying a product purchased by the customer, and a product price of the product purchased by the customer When,
Based on the set of customer purchase information acquired by the acquisition means, for each of the customers, product price information representing the product price of each of the products purchased by the customer, for each of the products, Data generation means for generating purchase price information representing each product price when the product is purchased, and purchase number information representing each product identification information of each of the products purchased by the customer for each of the customers; ,
For each of the customers, based on the product price information of the customer generated by the data generation means, the probability distribution of the product price is estimated as an internal reference price that is a price range that the customer considers purchasing. An internal reference price learning means to
Product price range learning means for estimating a probability distribution of the product price as the price range of the product based on the purchase price information of the product generated by the data generation means for each of the products,
For each of the combination of the customer and the product, a probability distribution representing the probability attributed to each topic based on the customer's interest, obtained based on the purchase quantity information generated by the data generation means, and each topic Customer interest learning means for estimating a probability that the customer of the combination will purchase the product based on a probability distribution representing a probability that the product in the topic will be purchased.
For each of the products, the degree of similarity between the probability distribution of the product price estimated as the internal reference price for the customer to be predicted and the probability distribution of the product price estimated as the purchase price information for the product Similarity calculation means for calculating;
For each of the products, based on the similarity to the product calculated by the similarity calculation means, and the probability of purchasing the product estimated for the combination of the product and the prediction target customer, A score calculation means for calculating a score indicating a possibility that the prediction target customer purchases the product;
A purchase forecasting device including A purchase prediction method in a purchase prediction apparatus including an acquisition means, a data generation means, an internal reference price learning means, a product price range learning means, a customer interest learning means, a similarity calculation means, and a score calculation means There,
A set of customer purchase information including customer identification information for identifying a customer by the acquisition means, product identification information for identifying a product purchased by the customer, and a product price of the product purchased by the customer Get
Product price information representing the product price of each of the products purchased by the customer for each of the customers based on the set of customer purchase information acquired by the acquisition unit by the data generating unit, the product Purchase price information representing each product price when each of the products is purchased, and purchase number information representing each product identification information of each of the products purchased by the customer for each of the customers. Generate
As the internal reference price that is a price range that the customer considers purchasing based on the product price information of the customer generated by the data generating means for each of the customers by the internal reference price learning means. , Estimate the probability distribution of the product price,
The product price range learning means estimates the probability distribution of the product price as the price range of the product based on the purchase price information of the product generated by the data generation means for each of the products,
The probability of belonging to each topic based on the customer's interest, determined based on the purchase quantity information generated by the data generation unit, for each combination of the customer and the product by the customer interest learning unit. Estimating the probability that the customer of the combination will purchase the product based on the probability distribution that represents and the probability distribution that represents the probability that the product on the topic will be purchased for each topic;
For each of the products, a probability distribution of the product price estimated as the internal reference price for each of the products, and the product price estimated as the purchase price information for the product. Calculate the similarity with the probability distribution,
Probability of purchasing the product estimated by the score calculation means for each of the products for the combination of the similarity calculated by the similarity calculation means with the product and the prediction target customer Based on the above, a purchase prediction method for calculating a score indicating a possibility that the prediction target customer purchases the product.   The program for functioning a computer as each means of the purchase prediction apparatus of Claim 1.

Images