JP2016212792A - Commodity recommendation device, commodity recommendation method and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To reduce a calculation amount required for recommending a commodity to a user.SOLUTION: A ranking acquisition part 161 of a ranking correction part 16 acquires ranking information corresponding to a customer ID being an object of correction of the ranking information, from a ranking DB 14. A purchase interval reflection part 162 corrects the rank value contained in the ranking information acquired by the ranking acquisition part 161, based on a commodity purchase interval. A store recommendation reflection part 163 corrects each rank value corrected by the purchase interval reflection part 162, based on recommendation of the store. A recommendation commodity presentation part 17 determines ranking of each commodity, based on the rank value corrected by the store recommendation reflection part 163, and presents the determined ranking and a commodity ID to a user of a communication terminal 3.SELECTED DRAWING: Figure 2

Description

  The present invention relates to a technique for recommending a product.

  Conventionally, collaborative filtering is known as one of product recommendation methods. Collaborative filtering is a method of recommending products purchased by other users who have purchased those products to users who are the product recommendation target, based on products previously purchased by the product recommendation target user. For example, in Patent Document 1, an adjacency matrix is generated from purchase history data of a customer, and the adjacency matrix is analyzed to acquire a time-series order relationship between products related to purchase. In this case, it is described that a product having a time-series order relationship with the product is recommended to the customer. In Patent Document 2, the purchase order relationship between products and the average value of the time difference between purchases are obtained based on the purchase history of the customer, and when a customer purchases a product, the order relationship with the product is obtained. It is described that a direct mail of a certain other product is sent at a delivery timing determined based on an average value of the time difference of purchase of these products.

JP 2002-279279 A Japanese Patent Laid-Open No. 9-305571

  As one of product recommendation methods using collaborative filtering, a method using RWR (Random Walk with Restart) is known. RWR is a method of calculating the degree of association between nodes in graph theory, and is particularly characterized in that the degree of association can be calculated based on the structural characteristics of the graph. The product recommendation method using RWR has a characteristic that it is more resistant to data sparsity than the method using machine learning, which is a representative example of collaborative filtering, and a method using singular value decomposition even in the method using the same matrix There is an advantage that the calculation is lighter than.

  However, the product recommendation method using RWR has a problem that it is unsuitable for a product recommendation system with a high update frequency because of the large number of customers and the number of recommended products. This is because every time the purchase history data of a product used for product ranking generation is updated, the calculation for constructing the graph structure of the product must be performed again. As the number of customers and the number of recommended products increase, the product graph increases, and it takes time to update the ranking in proportion to this.

  The present invention has been made in view of such circumstances, and an object thereof is to reduce the amount of calculation required when recommending a product to a user.

  In order to solve the above problems, the present invention provides a plurality of items associated with one user identification information from a product recommendation degree storage unit that stores user identification information, product identification information, and recommendation degree in association with each other. And a recommendation level acquisition unit that acquires a recommendation level associated with each of the plurality of product identification information, user identification information, product identification information, and use date and time are stored in association with each other. Each of the product usage history storage unit includes a first usage history acquisition unit that acquires a usage date and time associated with each of the plurality of product identification information items, and two product identification information items constituting the plurality of product identification information items. For the pair, the use interval from the use of the first product identified by the first product identification information to the use of the second product identified by the second product identification information is acquired. Based on the usage date A usage interval specifying unit to be specified; a recommendation level correcting unit that corrects the acquired recommendation level based on the specified usage interval; and the plurality of product identification information based on the corrected recommendation level There is provided a product recommendation device comprising a recommended product presentation unit that presents at least a part of the product to a user identified by the one user identification information.

  In a preferred aspect, the recommendation level correction unit determines a recommendation level corresponding to the second product identification information based on a use interval from when the first product is used to when the second product is used. To correct.

  In a further preferred aspect, the use interval specifying unit specifies a frequency distribution of use intervals from the use of the first product to the use of the second product, and the recommendation level correcting unit includes: The recommendation level corresponding to the second product identification information is corrected based on the frequency of the usage interval corresponding to the time interval that has passed since the first product was last used.

  In a further preferred aspect, the first usage history acquisition unit is associated with any one of the plurality of user identification information including the one user identification information and each of the plurality of product identification information. The date and time is acquired from the product use history storage unit.

  In a further preferred aspect, the apparatus further comprises a second usage history acquisition unit that acquires a usage date and time associated with the one user identification information and each of the plurality of product identification information from the product usage history storage unit, The usage interval specifying unit is a usage interval from when the first product is used until the second product is used for each set of two product identification information constituting the plurality of product identification information. Is determined based on the use date and time acquired by the first use history acquisition unit, and the first product is used for each set of two product identification information constituting the plurality of product identification information. A usage interval from when the second product is used until it is used is specified based on the usage date and time acquired by the second usage history acquisition unit.

  Further, the present invention is a product recommendation method executed by the product recommendation device, wherein the product recommendation level storage unit stores user identification information, product identification information, and recommendation level in association with each other. Acquiring a plurality of product identification information associated with the user identification information and a recommendation level associated with each of the plurality of product identification information, user identification information, product identification information, and use date and time. Each set consisting of a step of acquiring a use date and time associated with each of the plurality of product identification information, and two product identification information constituting the plurality of product identification information, from a product use history storage unit stored in association with each other The use interval from the use of the first product identified by the first product identification information to the use of the second product identified by the second product identification information is acquired. Profit Specifying based on date and time; correcting the acquired recommendation level based on the specified usage interval; and at least one of the plurality of product identification information based on the corrected recommendation level And a step of presenting a part to a user identified by the one user identification information.

  In addition, the present invention provides a plurality of product identifications associated with one user identification information from a product recommendation degree storage unit that stores user identification information, product identification information, and recommendation degree in association with each other on a computer. A product usage history storage unit that stores information and a recommendation level associated with each of the plurality of product identification information, user identification information, product identification information, and use date and time in association with each other. , A step of acquiring a use date and time associated with each of the plurality of product identification information and each set of two product identification information constituting the plurality of product identification information is identified by the first product identification information. A usage interval from when the first product is used until the second product identified by the second product identification information is used based on the acquired use date and time; Correcting the acquired recommendation level based on the specified usage interval; and identifying at least a part of the plurality of product identification information based on the corrected recommendation level as the one user identification A program for executing a step presented to a user identified by information is provided.

  According to the present invention, it is possible to reduce the amount of calculation required when recommending a product to a user.

It is a figure which shows an example of a structure of the goods recommendation system. It is a figure which shows an example of a structure of the goods recommendation apparatus. It is a figure which shows an example of purchase data DB12. It is a figure which shows an example of a graph structure. It is a figure which shows an example of ranking DB14. It is a figure which shows an example of store recommendation goods DB15. It is a figure which shows an example of the histogram which shows distribution of the frequency of the purchase interval of a customer's individual goods. It is a figure which shows an example of the value of the probability calculated from the frequency of the purchase interval of a customer individual product. It is a figure which shows an example of the graph of the value of the probability calculated from the frequency of the purchase interval of a customer individual goods. It is a figure which shows an example of the histogram which shows frequency distribution of the purchase interval of the goods of the whole customer. It is a figure which shows an example of the value of the probability calculated from the frequency of the purchase interval of the goods of the whole customer. It is a figure which shows an example of the graph of the value of the probability calculated from the frequency of the purchase interval of the goods of the whole customer. It is a figure which shows an example of addition of an individual space | interval score and a whole space | interval score. It is a figure which shows an example of a purchase space | interval score. It is a figure which shows an example of the graph of a purchase space | interval score. It is a flowchart which shows an example of a ranking production | generation process. It is a flowchart which shows an example of a ranking correction process. It is a flowchart which shows an example of a purchase space | interval reflection process. It is a flowchart which shows an example of an individual space | interval score calculation process. It is a flowchart which shows an example of a whole space | interval score calculation process.

1. Embodiment 1-1. Configuration of Product Recommendation System 100 FIG. 1 is a diagram illustrating an example of a configuration of a product recommendation system 100 according to an embodiment of the present invention. As shown in FIG. 1, the product recommendation system 100 includes a product recommendation device 1, a POS (Point of Sale) terminal 2, and a communication terminal 3. These devices are connected to each other via a communication line 4. The communication line 4 is configured by, for example, the Internet, a mobile phone network, or a wireless LAN (Local Area Network). Although only one POS terminal 2 and one communication terminal 3 are shown in FIG. 1 for convenience of explanation, the product recommendation system 100 may include a plurality of each.

  The product recommendation device 1 is a server device that collects POS purchase information from the POS terminal 2 and recommends products to customers who own the communication terminal 3 based on this information.

  The POS terminal 2 is a register machine corresponding to the POS system installed in a store such as a supermarket or a convenience store. The POS terminal 2 transmits POS purchase information to the product recommendation device 1. The POS purchase information includes a customer ID, purchase date and time, and a product ID.

  The communication terminal 3 is a computer device owned by a customer. Specifically, it is a portable electronic device such as a smartphone, a mobile phone, and a tablet terminal. Alternatively, it is a stationary electronic device such as a personal computer. The communication terminal 3 requests the recommended product information from the recommended product device 1 and displays the recommended product information received from the recommended product device 1 on the screen.

1-2. Configuration of Product Recommendation Device 1 FIG. 2 is a diagram illustrating an example of the configuration of the product recommendation device 1. As shown in FIG. 2, the product recommendation device 1 generally includes a purchase data acquisition unit 11, a purchase data database (hereinafter, “database” is referred to as “DB”) 12, a ranking generation unit 13, and a ranking DB 14. And a store recommended product DB 15, a ranking correction unit 16, and a recommended product presentation unit 17. These functions are realized, for example, by executing a program by an arithmetic device such as a CPU (Central Processing Unit).

  The purchase data acquisition unit 11 stores the POS purchase information transmitted from the POS terminal 2 in the purchase data DB 12. Specifically, the customer ID, the purchase date and time, and the product ID are associated with each other and stored in the purchase data DB 12.

  The purchase data DB 12 is a database that stores POS purchase information transmitted from the POS terminal 2. The purchase data DB 12 is an example of a “product use history storage unit” according to the present invention. FIG. 3 is a diagram illustrating an example of the purchase data DB 12. Each record constituting the purchase data DB 12 is composed of fields of customer ID, purchase date and time, and product ID. The customer ID is identification information for identifying a user of the communication terminal 3 who uses the store where the POS terminal 2 is installed. The product ID is identification information for identifying a product, and is, for example, a JAN (Japanese Article Number) code.

  The ranking generation unit 13 generates product ranking information based on the POS purchase information stored in the purchase data DB 12 and stores the generated ranking information in the ranking DB 14. The ranking generation unit 13 generates product ranking information for each customer based on POS purchase information within a predetermined period (for example, within the past year) at a predetermined period (for example, one month period). As illustrated in FIG. 2, the ranking generation unit 13 includes a graph construction unit 131, a weighting unit 132, and a rank value calculation unit 133.

  The graph construction unit 131 constructs a product graph structure based on the POS purchase information stored in the purchase data DB 12. Specifically, the purchase date and product ID corresponding to the customer for which ranking information is generated is acquired from the purchase data DB 12, and the graph structure of the product is constructed based on these information. At this time, a purchase date and time and a corresponding product ID within a predetermined period (for example, within the past year) are acquired. The constructed graph structure includes nodes representing products and bidirectional edges that link two products purchased on the same day.

  FIG. 4 is a diagram illustrating an example of a graph structure. The graph structure shown in FIG. 4 indicates that “bacon” and “ground meat”, “bacon” and “carrot”, “bacon” and “cabbage” were purchased on the same day. It also shows that “carrot” and “soy sauce”, “carrot” and “asparagus” were purchased on the same day.

  The weight assigning unit 132 assigns a weight to each edge of the graph structure constructed by the graph constructing unit 131. A transition probability value is assigned as the weight. Specifically, a value obtained by dividing the number of purchases of two products linked by an edge on the same day by the sum of the number of purchases of a transition destination product linked by the edge is given as a weight to the edge.

  In the graph structure shown in FIG. 4, as the weight of the edge between “bacon” and “minced meat”, the number of times “bacon” and “minced meat” were purchased on the same day “1 (times)” Number of purchases of “minced meat” “1 (times)” and “Bacon” Number of purchases of “carrot” on the same day “1 (times)” Number of purchases of “cabbage” on the same day as “Bacon” “1 ( The value “1/3” divided by the sum “3 (times)” is given. As another example, as the weight of the edge between “carrot” and “soy sauce”, the number of times “1 (time)” purchased “carrot” and “soy sauce” on the same day is changed to “soy sauce” on the same day as “carrot”. "1 (times)" and "carrots" and the number of times "asparagus" purchased on the same day "1 (times)" divided by the sum "2 (times)" is "1/2" Has been granted.

  The rank value calculating unit 133 calculates the rank value of each product in the graph structure in which the weight is given to the edge by the weight assigning unit 132. At this time, the rank value is calculated using RWR. Here, RWR is an algorithm for calculating the relevance (or similarity) of each node adjacent to the node that is the starting point on the graph structure. This degree of association corresponds to the rank value in the present embodiment. In RWR, a random walk is started from an origin node a, moves randomly to an adjacent node with a probability proportional to the edge weight, and returns to the origin node a with a certain probability every time the node is reached. The relevance is a steady state probability obtained at each node as a result of recursively repeating this operation. That is, the relevance of the node b viewed from the node a on the graph structure is calculated as the steady state probability of the node b obtained by random walking from the node a. Regarding the calculation method of the relevance using this RWR, for example, “M. Gori, 1 other person,“ ItemRank: A Random-Walk Based Scoring Algorithm for Recommender Engines ”, Proc. IJCAI '07, p.2766-2771 "It is described in.

  In the graph structure shown in FIG. 4, the rank value (importance in FIG. 4) of “bacon” that is the starting node is “1.0 × 0.9 (= 0.9)”. As another example, the rank value (importance in FIG. 4) of “minced meat” when the number of recursively repeated is 3 is “1.0 × 0.9 × (1/3) (= 0.3)”. is there. The number of times of recursively repeating is not limited to three, and the steady state probability obtained by converging the value is used as the rank value.

  When the rank value calculation unit 133 calculates the rank value of each product, the ranking generation unit 13 associates the rank value of a predetermined number of products (for example, products with rank values within the top 20) with the product ID. In addition, it is stored in the ranking DB 14. At this time, the customer ID, the previous purchase date and time of the product, and the ranking of the product are stored in the ranking DB 14 in association with the rank value and the product ID. In addition, when the most recently purchased product is included in the upper predetermined number of products, the product may be excluded and the lower product may be moved up. This is because the rank value is calculated high for the latest purchased product.

  Next, the ranking DB 14 is a database that stores ranking information generated by the ranking generation unit 13. The ranking DB 14 is an example of a “product recommendation degree storage unit” according to the present invention. FIG. 5 is a diagram illustrating an example of the ranking DB 14. Each record constituting the ranking DB 14 is configured by customer ID, product ID, previous purchase date / time, rank value, and ranking fields. Here, the previous purchase date and time is the date and time when the product identified by the corresponding product ID was recently purchased (in other words, the date and time when the product was last purchased). The rank value is a value indicating the recommendation level of the product, and is calculated by the rank value calculation unit 133. The ranking is the ranking of products determined based on the corresponding rank value. The smaller the corresponding rank value is, the smaller the value is given. The ranking DB 14 stores, for example, information on products within the 20th ranking for each customer.

  The store recommended product DB 15 is a database that stores information on products to be recommended to customers at the store where the POS terminal 2 is installed. Specifically, it is a database that stores correction values used for correcting the rank value of store recommended products. FIG. 6 is a diagram illustrating an example of the store recommended product DB 15. Each record that configures the store recommended product DB 15 includes fields of a product ID, a recommendation start date / time, a recommendation end date / time, and a correction value. Here, the recommendation start date and time and the recommendation end date and time are respectively the start date and time and end date and time of the period in which the corresponding correction value is added to the rank value of the product identified by the corresponding product ID. The correction value is a value added to the rank value of the product identified by the corresponding product ID.

  The information stored in the store recommended product DB 15 is updated, for example, by the manager of the store where the POS terminal 2 is installed via the communication line 4 using a communication device (not shown). The store manager determines a recommended product or a non-recommended product according to the location, season, or stock of the store, for example. A positive value is set as a correction value for recommended products, and a negative value is set as a correction value for non-recommended products.

  The ranking correction unit 16 corrects the ranking information stored in the ranking DB 14. Specifically, correction is performed to reflect the purchase interval of the product and the recommendation of the store in the rank value included in the ranking information. The ranking correction unit 16 corrects the ranking information at a frequency higher than the ranking generation by the ranking generation unit 13. For example, if ranking generation is performed on a monthly basis, the ranking information is corrected on a daily basis. Alternatively, it may be executed each time a user of the communication terminal 3 requests provision of recommended product information. As shown in FIG. 2, the ranking correction unit 16 is generally configured by a ranking acquisition unit 161, a purchase interval reflection unit 162, and a store recommendation reflection unit 163.

  The ranking acquisition unit 161 acquires the ranking information corresponding to the customer ID for which the ranking information is to be corrected from the ranking DB 14. The ranking acquisition unit 161 is an example of the “recommendation degree acquisition unit” according to the present invention. The customer ID may be acquired from the communication terminal 3 together with a request for providing recommended product information, or may be stored in advance in the product recommendation device 1 as a target for periodically providing recommended product information. The ranking acquisition unit 161 may acquire all the ranking information corresponding to the customer ID from the ranking DB 14, or correct only a predetermined upper number of ranking information (for example, ranking information having rank values within the top 20). You may make it acquire as.

  The purchase interval reflection unit 162 corrects the rank value included in the ranking information acquired by the ranking acquisition unit 161 based on the purchase interval of the product. Specifically, the rank value is corrected based on the purchase interval of the individual customer's product, which is the target of the correction of the ranking information, and the purchase interval of the entire customer's product registered in the purchase data DB 12. As shown in FIG. 2, the purchase interval reflecting unit 162 is roughly configured by an individual interval score calculating unit 1621, an overall interval score calculating unit 1622, an interval score calculating unit 1623, and a rank value correcting unit 1624. .

  The individual interval score calculation unit 1621 identifies the frequency distribution of the purchase interval of the customer's individual product for which the ranking information is to be corrected, and calculates a probability value based on each frequency. Here, the product purchase interval is a time interval for purchasing two different types of products. For example, it is the time interval from the purchase of the product A to the purchase of the product B at a later date. As will be described later, the product purchase interval may be a time interval for purchasing two products identified by the same product ID. The personal interval score calculation unit 1621 includes a histogram creation unit 16211 and a probability calculation unit 16212.

  The histogram creation unit 16211 creates a histogram indicating the frequency distribution of the purchase intervals of individual customer products for which the ranking information is to be corrected. The histogram creation unit 16211 is an example of the “second usage history acquisition unit” and the “use interval identification unit” according to the present invention. Specifically, the histogram creation unit 16211 first acquires the purchase date and time corresponding to the customer ID and each product ID included in the ranking information acquired by the ranking acquisition unit 161 from the purchase data DB 12. At this time, the purchase date and time within a predetermined period (for example, within the past year) is acquired. In addition, the histogram creation unit 16211 acquires the latest purchase date and time corresponding to the customer ID for which the ranking information is to be corrected and the product ID corresponding to this purchase date and time from the purchase data DB 12. Here, the latest purchase date and time may be the most recent purchase date and time, or may be one or more purchase dates and times belonging to a predetermined period from the time of this histogram creation process. And about the arbitrary combination of the product ID included in the ranking information acquired by the ranking acquisition unit 161 and the product ID corresponding to the latest purchase date and time, after purchasing the first product, until the second product is purchased. A histogram showing the frequency distribution of purchase intervals is created. For example, if the first product was purchased on “2015/4/1” and “2015/4/8”, the second product was purchased on “2015/4/3” and “2015/4/14” If the first product was purchased on “April 1, 2015” and then the second product was purchased on “April 3, 2015”, the purchase interval “2 days” The frequency increases by “1”. In addition, since the first product was purchased on “April 8, 2015” and the second product was purchased on “April 14, 2015”, the frequency of the purchase interval “6 days” is “ 1 ”increase. If the ranking information includes, for example, 20 product IDs and the number of product IDs corresponding to the most recent purchase date and time is 10, the histogram creation unit 16211 displays 200 (= 20 × 10). A histogram is created for each customer ID.

  FIG. 7 is a diagram illustrating an example of a histogram created by the histogram creation unit 16211. The horizontal axis of the histogram shown in FIG. 7 indicates the purchase interval (unit is week), and the vertical axis indicates the number of purchases. This histogram shows the frequency distribution of purchase intervals after purchasing the product A until purchasing the product B. This histogram shows, for example, that the product B is purchased once in the second week (between the 8th and 14th days) after the purchase of the product A. As another example, after the purchase of the product A, the product B is purchased twice in the fifth week (between the 29th and 35th days).

  The probability calculation unit 16212 calculates a probability value (individual interval score) based on each frequency of the histogram created by the histogram creation unit 16211. Specifically, each frequency is input to the kernel density function to calculate a probability value. The kernel density function p (x) is expressed as follows, for example.

ここで、Ｎは、処理対象の購入間隔の数であり、ｘ_iは、各購入間隔の度数である。バンド幅ｄは、例えば以下のように表される。

Here, N is the number of purchase intervals to be processed, and x _i is the frequency of each purchase interval. The bandwidth d is expressed as follows, for example.

ここで、σは、処理対象の購入間隔の度数についての標準偏差である。

Here, σ is a standard deviation for the frequency of the purchase interval to be processed.

  FIG. 8 is a diagram illustrating an example of probability values calculated by the probability calculation unit 16212. FIG. 9 is a diagram illustrating a graph of probability values. The horizontal axis of the graph shown in FIG. 9 indicates the purchase interval (unit is week), the left vertical axis indicates the number of purchases, and the right vertical axis indicates the probability. The probability calculation unit 16212 calculates a probability value for the frequency of each histogram created by the histogram creation unit 16211.

  Next, the overall interval score calculation unit 1622 specifies the distribution of the purchase intervals of the products for all customers registered in the purchase data DB 12, and calculates a probability value based on each frequency. Here, the product purchase interval is a time interval for purchasing two different types of products. For example, it is the time interval from the purchase of the product A to the purchase of the product B at a later date. As will be described later, the product purchase interval may be a time interval for purchasing two products identified by the same product ID. The overall interval score calculation unit 1622 is different from the individual interval score calculation unit 1621 in that probability calculation is performed based on the frequency of purchase intervals of the products of the entire customer, not a specific customer individual. The overall interval score calculation unit 1622 includes a histogram creation unit 16221 and a probability calculation unit 16222.

  The histogram creation unit 16221 creates a histogram indicating the frequency distribution of purchase intervals of products for all customers registered in the purchase data DB 12. The histogram creation unit 16221 is an example of the “first usage history acquisition unit” and the “usage interval identification unit” according to the present invention. Specifically, the histogram creation unit 16221 first acquires the purchase date and time corresponding to each product ID included in the ranking information acquired by the ranking acquisition unit 161 from the purchase data DB 12. At this time, the purchase date and time within a predetermined period (for example, within the past year) is acquired. Also, the histogram creation unit 16221 acquires the latest purchase date and time corresponding to all customer IDs and the product ID corresponding to this purchase date and time from the purchase data DB 12. Here, the latest purchase date and time may be the most recent purchase date and time, or may be one or more purchase dates and times belonging to a predetermined period from the time of this histogram creation process. The histogram creation unit 16221 then selects the second combination of the product ID included in the ranking information acquired by the ranking acquisition unit 161 and the product ID corresponding to the latest purchase date and time after purchasing the first product. A histogram showing the frequency distribution of purchase intervals until the product is purchased is created. For example, if the first product was purchased on “2015/4/1” and “2015/4/8”, the second product was purchased on “2015/4/3” and “2015/4/14” If the first product was purchased on “April 1, 2015” and then the second product was purchased on “April 3, 2015”, the purchase interval “2 days” The frequency increases by “1”. In addition, since the first product was purchased on “April 8, 2015” and the second product was purchased on “April 14, 2015”, the frequency of the purchase interval “6 days” is “ 1 ”increase. For example, when the ranking information includes 50 product IDs and the number of product IDs corresponding to the most recent purchase date / time is 50, the histogram creation unit 16221 has 2500 (= 50 × 50). Create a histogram. If the number of product IDs corresponding to the most recent purchase date / time is enormous and the number of histograms to be created is large and calculation is difficult, the product ID to be used is the product ID of the top 20% of products that are often purchased. You may make it limit to.

  FIG. 10 is a diagram illustrating an example of a histogram created by the histogram creation unit 16221. The horizontal axis of the histogram shown in FIG. 10 indicates the purchase interval (unit is week), and the vertical axis indicates the number of purchases (more specifically, the cumulative value of the number of purchases for each customer). This histogram shows the frequency distribution of purchase intervals after purchasing the product A until purchasing the product B. This histogram shows, for example, that the product B has been purchased 29 times in the first week (between the first day and the seventh day) after the purchase of the product A. As another example, after the purchase of the product A, the product B is purchased 14 times in the second week (between the 8th and 14th days).

  The probability calculation unit 16222 calculates a probability value (overall interval score) based on each frequency of the histogram created by the histogram creation unit 16211. Specifically, each frequency is input to the above-described kernel density function to calculate a probability value.

  FIG. 11 is a diagram illustrating an example of probability values calculated by the probability calculation unit 16222. FIG. 12 is a diagram illustrating a graph of probability values. The horizontal axis of the graph shown in FIG. 12 indicates the purchase interval (unit is a week), the vertical axis on the left side indicates the number of purchases (more specifically, the cumulative value of the number of purchases for each customer), and the vertical axis on the right side. The axis shows the probability. The probability calculation unit 16222 calculates a probability value for the frequency of each histogram created by the histogram creation unit 16221.

  Next, the interval score calculation unit 1623 has a probability value calculated by the individual interval score calculation unit 1621 based on each frequency of the histogram, and a probability calculated by the overall interval score calculation unit 1622 based on each frequency of the histogram. The purchase interval score is calculated by adding the value and dividing the obtained value by “2”. At this time, the interval score calculation unit 1623 adds two probability values for each purchase interval. In addition, the interval score calculation unit 1623 performs this addition and division processing between pairs of corresponding probability values. Specifically, it is performed between sets of probability values calculated for combinations of the same product that have the same purchase order. For example, after purchasing the product A, after purchasing the product A, the product B is purchased after purchasing the product A and the set of probability values calculated by the individual interval score calculation unit 1621 based on the frequency of purchase intervals until the product B is purchased. This is performed between a set of probability values calculated by the overall interval score calculation unit 1622 based on the frequency of the purchase interval until the purchase.

  FIG. 13 is a diagram illustrating an example of addition performed by the interval score calculation unit 1623. In the example of addition shown in FIG. 13, the upper set of probability values is a set of probability values calculated by the individual interval score calculation unit 1621 (see FIG. 8). That is, it is a set of probability values calculated on the basis of the frequency of purchase intervals of individual customer products for which the ranking information is to be corrected. This set of probability values is a set of probability values calculated based on the frequency of purchase intervals after purchasing the product A until the product B is purchased. The set of probability values in the middle row is a set of probability values calculated by the overall interval score calculation unit 1622 (see FIG. 11). That is, it is a set of probability values calculated based on the frequency of purchase intervals for the products of the entire customer. This set of probability values is also a set of probability values calculated based on the frequency of purchase intervals after purchasing the product A until the product B is purchased. The lower probability value set is a value set calculated by adding the upper probability value and the lower probability value for each purchase interval.

  In the example of addition illustrated in FIG. 13, for example, the value “0.156632” is calculated by adding the upper probability value “0.048172” and the middle probability value “0.10846” of the first week. Is shown. As another example, the value “0.212589” is calculated by adding the value “0.096845” of the upper stage probability of the second week and the value “0.115745” of the middle stage.

  FIG. 14 is a diagram illustrating an example of the purchase interval score finally obtained by the interval score calculation unit 1623. FIG. 15 is a diagram illustrating a graph of the purchase interval score. The horizontal axis of the graph shown in FIG. 15 indicates the purchase interval (unit is week), and the vertical axis indicates the probability. Each purchase interval score shown in FIG. 14 is a value obtained by dividing the probability value shown in the lower part of FIG. 13 by “2”. According to the graph shown in FIG. 15, the purchase interval score peaks in the fifth week and the ninth week, similarly to the probability value calculated by the individual interval score calculation unit 1621. Further, the purchase interval score is affected by the probability value calculated by the overall interval score calculation unit 1622 in the first week and the second week, and is calculated by the individual interval score calculation unit 1621. The value is larger than. In other words, the purchase interval score reflects both the purchase interval of the individual product of the customer and the purchase interval of the product for the entire customer.

  The rank value correction unit 1624 corrects the rank value included in the ranking information acquired by the ranking acquisition unit 161 based on the purchase interval score calculated by the interval score calculation unit 1623. The rank value correction unit 1624 is an example of the “recommendation degree correction unit” according to the present invention. The rank value correction unit 1624 performs the correction based on each set of purchase interval scores calculated by the interval score calculation unit 1623.

  Specifically, the rank value correction unit 1624 first determines the time interval from the last purchase of the preceding purchase product at the purchase interval of the product referred to in order to calculate the purchase interval score to the time of the correction process. Is identified. For example, when the rank value is corrected based on the purchase interval score calculated based on the frequency of the purchase interval from the purchase of the product A until the purchase of the product B, the date and time when the product A was recently purchased (in other words, And the time interval from the last purchase date and time to the time of the correction process. Here, the date and time when the product A was recently purchased is specified by referring to the previous purchase date and time associated with the product A in the ranking DB 14, for example.

  Next, the rank value correction unit 1624 specifies a purchase interval score corresponding to the specified time interval. For example, if the specified time interval is “3 days” and the purchase interval score to be referenced is as shown in FIG. 14, the purchase interval score “0.078316” for the first week is specified. As another example, if the specified time interval is “9 days”, the purchase interval score “0.106295” of the second week shown in FIG. 14 is specified.

  Finally, the rank value correcting unit 1624 multiplies the rank value of the subsequent purchase product in the purchase interval of the product referred to in order to calculate the purchase interval score by the specified purchase interval score. For example, when the rank value is corrected based on the purchase interval score calculated based on the frequency of the purchase interval from the purchase of the product A until the purchase of the product B, the rank value of the product B is specified. Multiply by purchase interval score.

  Next, the store recommendation reflecting unit 163 corrects each rank value corrected by the purchase interval reflecting unit 162 based on the recommendation of the store. Specifically, each rank value is corrected based on information stored in the store recommended product DB 15. For example, in the store recommended product DB 15 shown in FIG. 6, since the correction value “+0.1” is set for the product ID “11111111”, the execution date and time of the correction process is the recommended start date “2015/5/1”. And the recommendation end date “2015/5/31”, “0.1” is added to the rank value corresponding to the product ID “11111111”.

  Next, the recommended product presentation unit 17 determines the ranking of each product based on the rank value corrected by the ranking correction unit 16, and presents the determined ranking to the user of the communication terminal 3 together with the product ID. At this time, the recommended product presentation unit 17 gives a ranking of a smaller value to the product as the corresponding rank value is larger. The recommended product presentation unit 17 specifically transmits screen data for displaying the product ID and the ranking to the communication terminal 3.

  The recommended product presentation unit 17 presents the product ID and the ranking to the user of the communication terminal 3 only for a predetermined number of products (for example, products within the top 5 ranking) based on the determined ranking. You may do it. The recommended product presentation unit 17 may present a product name, a product image, and a product price instead of or in addition to the product ID. In addition, when the product recommendation device 1 includes a display device, the recommended product presentation unit 17 displays the above screen data on the display device without transmitting it to the communication terminal 3, thereby displaying the communication terminal 3. The product ID and the ranking may be presented to the user.

1-3. Operation The operation of the product recommendation device 1 will be described. Specifically, ranking generation processing and ranking correction processing will be described.

1-3-1. Ranking Generation Processing FIG. 16 is a flowchart illustrating an example of ranking generation processing. The following description assumes the case where ranking information for one customer is generated. The ranking generation process is executed by the ranking generation unit 13, for example, at a predetermined cycle (for example, one month cycle).

  In step Sa1 of this process, the graph construction unit 131 constructs a graph structure of the product based on the POS purchase information stored in the purchase data DB 12. Specifically, the purchase date and product ID corresponding to the customer for which ranking information is generated is acquired from the purchase data DB 12, and the graph structure of the product is constructed based on these information. The constructed graph structure includes nodes representing products and bidirectional edges that link two products purchased on the same day.

  Next, the weight assigning unit 132 assigns a weight to each edge of the graph structure constructed in Step Sa1 (Step Sa2). Specifically, a value obtained by dividing the number of purchases of two products linked by an edge on the same day by the number of purchases of a transition destination product is assigned to the edge as a weight.

  Next, the rank value calculation unit 133 calculates the rank value of each product in the graph structure in which the edge is weighted in step Sa2 (step Sa3). At this time, the rank value is calculated using the above-mentioned RWR.

Next, the ranking generation unit 13 stores the rank value of each product calculated in step Sa3 in the ranking DB 14 in association with the product ID (step Sa4). At this time, only the rank value of the upper predetermined number of products (for example, products having rank values within the top 20) may be stored in the ranking DB 14 in association with the product ID. When storing the rank value and the product ID, the customer ID, the previous purchase date and time of the product, and the ranking of the product are stored in the ranking DB 14 in association with them.
This completes the description of the ranking generation process.

1-3-2. Ranking Correction Process FIG. 17 is a flowchart showing an example of the ranking correction process. The ranking correction process is executed by the ranking correction unit 16 at a higher frequency than the ranking generation process. The ranking correction process is executed, for example, when a user of the communication terminal 3 requests provision of recommended product information.

  In step Sb1 of this process, the ranking acquisition unit 161 acquires ranking information corresponding to the customer ID for which the ranking information is to be corrected from the ranking DB. The customer ID is acquired from the communication terminal 3 together with a request for providing information on recommended products, for example. Here, the ranking acquisition unit 161 may acquire only a predetermined upper number of ranking information (for example, ranking information having rank values within the top 20) from the ranking DB 14 as correction targets.

  Next, the purchase interval reflecting unit 162 corrects the rank value included in the ranking information acquired in step Sb1 based on the purchase interval of the product (step Sb2). Specifically, the rank value is corrected based on the purchase interval of the individual customer's product, which is the target of the correction of the ranking information, and the purchase interval of the entire customer's product registered in the purchase data DB 12.

  FIG. 18 is a flowchart illustrating an example of the purchase interval reflecting process executed in step Sb2. In step Sc1 of this process, the individual interval score calculation unit 1621 specifies the frequency distribution of the purchase interval of the customer's individual product for which the ranking information is to be corrected, and calculates the probability value based on each frequency. Here, the product purchase interval is a time interval for purchasing two different types of products. As will be described later, the product purchase interval may be a time interval for purchasing two products identified by the same product ID.

  FIG. 19 is a flowchart showing an example of the personal interval score calculation process executed in step Sc1. In step Sd1 of this process, the histogram creation unit 16211 creates a histogram indicating the distribution of frequency of purchase intervals of individual customer products that are targets of ranking information correction. Specifically, first, the purchase date and time corresponding to the customer ID and each product ID included in the ranking information acquired in step Sb1 is acquired from the purchase data DB 12. Further, the latest purchase date and time corresponding to the customer ID for which the ranking information is to be corrected and the product ID corresponding to the purchase date and time are acquired from the purchase data DB 12. Then, for any combination of the product ID included in the ranking information and the product ID corresponding to the latest purchase date and time, the frequency of the purchase interval until the purchase of the second product after the purchase of the first product. Create a histogram showing the distribution.

  Next, the probability calculation unit 16212 calculates a probability value based on each frequency of the histogram created in step Sd1 (step Sd2). Specifically, each frequency is input to the kernel density function to calculate a probability value. The probability calculation unit 16212 calculates a probability value for the frequency of each histogram created in step Sd1.

  In step Sc2 of FIG. 18, the overall interval score calculation unit 1622 specifies the distribution of purchase intervals of products for all customers registered in the purchase data DB 12, and calculates a probability value based on each frequency. Here, the product purchase interval is a time interval for purchasing two different types of products. As will be described later, the product purchase interval may be a time interval for purchasing two products identified by the same product ID.

  FIG. 20 is a flowchart illustrating an example of the overall interval score calculation process executed in step Sc2. In step Se <b> 1 of this process, the histogram creation unit 16221 creates a histogram indicating the frequency distribution of purchase intervals of products for all customers registered in the purchase data DB 12. Specifically, first, the purchase date and time corresponding to each product ID included in the ranking information acquired in step Sb1 is acquired from the purchase data DB 12. Further, the latest purchase date and time corresponding to all customer IDs and the product ID corresponding to this purchase date and time are acquired from the purchase data DB 12. Then, for any combination of the product ID included in the ranking information and the product ID corresponding to the latest purchase date and time, the frequency of the purchase interval until the purchase of the second product after the purchase of the first product. Create a histogram showing the distribution.

  The probability calculation unit 16222 calculates a probability value based on each frequency of the histogram created in step Se1 (step Se2). Specifically, each frequency is input to the above-described kernel density function to calculate a probability value. The probability calculation unit 16222 calculates a probability value for the frequency of each histogram created in step Se1.

  In step Sc3 in FIG. 18, the interval score calculation unit 1623 adds the probability value calculated in step Sc1 and the probability value calculated in step Sc2, and divides the obtained value by “2”. By doing so, a purchase interval score is calculated. At this time, the interval score calculation unit 1623 adds two probability values for each purchase interval. In addition, the interval score calculation unit 1623 performs this addition and division processing between pairs of corresponding probability values.

  Next, the rank value correcting unit 1624 corrects the rank value included in the ranking information acquired in step Sb1 based on the purchase interval score calculated in step Sc3 (step Sc4). The rank value correction unit 1624 performs the correction based on each set of purchase interval scores calculated in step Sc3.

  Specifically, the rank value correction unit 1624 first specifies the time interval from the last purchase of the pre-purchased product in the purchase interval of the product referred to in order to calculate the purchase interval score to the time of the correction process. To do. Next, the rank value correction unit 1624 specifies a purchase interval score corresponding to the specified time interval. Finally, the rank value correcting unit 1624 multiplies the rank value of the subsequent purchase product in the purchase interval of the product referred to in order to calculate the purchase interval score by the specified purchase interval score.

  Next, in step Sb3 of FIG. 17, the store recommendation reflecting unit 163 corrects each rank value corrected in step Sb2 based on the recommendation of the store. Specifically, each rank value is corrected based on information stored in the store recommended product DB 15.

Next, the recommended product presentation unit 17 determines the ranking of each product based on the rank value corrected in step Sb3, and presents the ranking determined together with the product ID to the user of the communication terminal 3 (step Sb4). Specifically, screen data for displaying the product ID and the ranking is transmitted to the communication terminal 3.
This completes the description of the ranking correction process.

  According to the product recommendation device 1 according to the present embodiment described above, the ranking generation that requires recalculation using the RWR is not executed when the recommended product information is provided to the customer, and only the ranking correction is performed. Therefore, the amount of calculation required for recommending a product to a customer can be reduced. For this reason, the product recommendation device 1 according to the present embodiment is compatible with a product recommendation system with a large number of customers and recommended product items and a high update frequency.

  In addition, according to the product recommendation device 1 according to the present embodiment, the ranking is corrected based on the purchase interval of each customer's product, so it is possible to recommend the product at a timing when the customer is highly likely to purchase. it can. In addition, since the ranking is corrected based on the purchase interval of each customer's product as well as the purchase interval of each customer's product, even if there is a lack of data about the purchase interval of individual customer products, This can be supplemented by data on the purchase intervals of the products of the entire customer.

  Further, according to the product recommendation device 1 according to the present embodiment, since the ranking is corrected based on the recommendation of the store, the product can be recommended to the customer according to the location, season, and stock of the store.

2. Modifications The above embodiment may be modified as shown below. Further, the following modifications may be combined with each other.

2-1. Modification 1
In the description of the above embodiment, food is given as an example of the product (see, for example, FIG. 4), but this is only an example of the product. In this specification, a product is a good or service that is the object of commercial transactions. Here, the goods may be tangible or intangible (for example, digital content). Further, the commercial transaction includes not only buying and selling but also renting (rental). That is, the product recommended for the customer in the above embodiment may be a rental target.

2-2. Modification 2
In the ranking generation process according to the above embodiment, a product recommendation method using RWR is used, but this is only an example of a ranking generation method. In the above ranking generation process, instead of the product recommendation method using RWR, other conventional collaborative filtering (for example, “Sarwar, 3 others,“ Item-based Collaborative Filtering Recommendation Algorithms ”, Proc. WWW '01, p.285-295 ") may be used to generate ranking information.

2-3. Modification 3
In the ranking correction processing according to the above embodiment, the kernel density function is used to calculate the probability value based on the frequency of the histogram, but this is only an example of a means for calculating the probability value. . In the above ranking correction process, the interval score may be calculated by applying the frequency of the histogram to an exponential distribution or a normal distribution instead of the kernel density function.

2-4. Modification 4
In the ranking correction processing according to the above embodiment, when correcting the rank value based on the purchase interval of the product, the time interval for purchasing two different types of products is referred to. You may make it refer to the time interval which purchases goods. Specifically, the rank value may be corrected based on the time interval from the purchase of the product A to the next purchase of the product A.

  In this case, in the above-described embodiment, the histogram creation unit 16211 and the histogram creation unit 16221 purchase the product for each product ID included in the ranking information acquired by the ranking acquisition unit 161, and then select the product. Create a histogram showing the frequency distribution of purchase intervals until purchase. The interval score calculation unit 1623 adds the value of the probability calculated by the individual interval score calculation unit 1621 and the value of the probability calculated by the overall interval score calculation unit 1622 for the same type of product. Is divided by “2” to calculate the purchase interval score. The rank value correction unit 1624 specifies a time interval from the last purchase of a product corresponding to the purchase interval score to the time of the correction process, specifies a purchase interval score corresponding to the specified time interval, and Multiply the product rank by the specified purchase interval score.

  Also according to this modification, it is possible to recommend a product at a timing when the customer is highly likely to purchase.

2-5. Modification 5
In the ranking correction process according to the above embodiment, the rank value correction process based on the store recommendation by the store recommendation reflection unit 163 may be omitted. Further, in the ranking correction process according to the above-described embodiment, the purchase interval reflecting unit 162 corrects the rank value based on only one of the customer's individual product purchase interval and the customer's entire product purchase interval. You may make it do. That is, the rank value correction unit 1624 is configured to determine the rank value based on only one of the probability value calculated by the individual interval score calculation unit 1621 and the probability value calculated by the overall interval score calculation unit 1622. May be corrected.

2-6. Modification 6
In the above embodiment, the product recommendation device 1 includes the purchase data acquisition unit 11, the purchase data DB 12, the ranking generation unit 13, the ranking DB 14, and the store recommended product DB 15. It may be provided in another device connected to the product recommendation device 1 via the communication line 4.

2-7. Modification 7
The program executed in the product recommendation device 1 according to the above embodiment may be provided via a recording medium readable by a computer device. Here, the recording medium is, for example, a magnetic recording medium such as a magnetic tape or a magnetic disk, an optical recording medium such as an optical disk, a magneto-optical recording medium, or a semiconductor memory. The program may be distributed via a network such as the Internet.

DESCRIPTION OF SYMBOLS 1 ... Product recommendation apparatus, 2 ... POS terminal, 3 ... Communication terminal, 4 ... Communication line, 11 ... Purchase data acquisition part, 12 ... Purchase data DB, 13 ... Ranking generation part, 14 ... Ranking DB, 15 ... Store recommendation product DB: 16 ... Ranking correction unit, 17 ... Recommended product presentation unit, 100 ... Product recommendation system, 131 ... Graph construction unit, 132 ... Weight assignment unit, 133 ... Rank value calculation unit, 161 ... Ranking acquisition unit, 162 ... Purchase interval Reflection unit, 163 ... store recommendation reflection unit, 1621 ... personal interval score calculation unit, 1622 ... overall interval score calculation unit, 1623 ... interval score calculation unit, 1624 ... rank value correction unit, 16211 ... histogram creation unit, 16212 ... probability calculation Part, 16221 ... histogram creation part, 16222 ... probability calculation part

Claims (7)

From a product recommendation degree storage unit that stores user identification information, product identification information, and recommendation level in association with each other, a plurality of product identification information associated with one user identification information, and the plurality of product identification information A recommendation level acquisition unit that acquires a recommendation level associated with each of the
First usage history acquisition for acquiring a usage date and time associated with each of the plurality of product identification information from a product usage history storage unit that stores user identification information, product identification information, and usage date and time And
For each set of two product identification information constituting the plurality of product identification information, the first product identified by the first product identification information is used and then identified by the second product identification information. A use interval specifying unit for specifying a use interval until the second product is used based on the acquired use date and time;
A recommendation level correction unit that corrects the acquired recommendation level based on the specified use interval;
A product recommendation device comprising: a recommended product presentation unit that presents at least a part of the plurality of product identification information to a user identified by the one user identification information based on the corrected recommendation degree .   The recommendation level correction unit corrects the recommendation level corresponding to the second product identification information based on a use interval from when the first product is used until the second product is used. The product recommendation device according to claim 1. The usage interval specifying unit specifies a frequency distribution of usage intervals from when the first product is used to when the second product is used,
The recommendation level correction unit determines a recommendation level corresponding to the second product identification information based on a frequency of use intervals corresponding to a time interval that has passed since the first product was last used. The product recommendation device according to claim 2, wherein correction is performed.   The first use history acquisition unit sets a use date and time associated with any one of the plurality of user identification information including the one user identification information and each of the plurality of product identification information. The product recommendation device according to claim 1, wherein the product recommendation device is acquired from a history storage unit. A second usage history acquisition unit that acquires from the product usage history storage unit a usage date and time associated with the one user identification information and each of the plurality of product identification information;
The usage interval specifying unit is a usage interval from when the first product is used until the second product is used for each set of two product identification information constituting the plurality of product identification information. Is determined based on the use date and time acquired by the first use history acquisition unit, and the first product is used for each set of two product identification information constituting the plurality of product identification information. 5. The product recommendation according to claim 4, wherein a use interval from when the second product is used until the second product is used is specified based on a use date and time acquired by the second use history acquisition unit. apparatus. A product recommendation method executed by a product recommendation device,
From a product recommendation degree storage unit that stores user identification information, product identification information, and recommendation level in association with each other, a plurality of product identification information associated with one user identification information, and the plurality of product identification information Obtaining a recommendation level associated with each of the
Obtaining a use date and time associated with each of the plurality of product identification information from a product use history storage unit that stores user identification information, product identification information, and use date and time;
For each set of two product identification information constituting the plurality of product identification information, the first product identified by the first product identification information is used and then identified by the second product identification information. Specifying a use interval until the second product is used based on the acquired use date and time;
Correcting the acquired recommendation degree based on the specified use interval;
Presenting at least a part of the plurality of product identification information to a user identified by the one user identification information based on the corrected recommendation degree. On the computer,
From a product recommendation degree storage unit that stores user identification information, product identification information, and recommendation level in association with each other, a plurality of product identification information associated with one user identification information, and the plurality of product identification information Obtaining a recommendation level associated with each of the
Obtaining a use date and time associated with each of the plurality of product identification information from a product use history storage unit that stores user identification information, product identification information, and use date and time;
For each set of two product identification information constituting the plurality of product identification information, the first product identified by the first product identification information is used and then identified by the second product identification information. Specifying a use interval until the second product is used based on the acquired use date and time;
Correcting the acquired recommendation degree based on the specified use interval;
Presenting at least a part of the plurality of product identification information to the user identified by the one user identification information based on the corrected recommendation degree.

Images