JP2010134733A - Information recommendation device, information recommendation method, and program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide an information recommendation device or the like which recommends proper information to a user in consideration of various factors such as attributes and interests of a target user and biases of various factors caused by an object to be recommended. <P>SOLUTION: A CPU of an information recommendation device 5 associates articles with a plurality of factors caused by users or the articles and calculates event probabilities (factor values) of factor events of the factors to generate an article analysis result TB 511. The CPU calculates biases A(P)i of the factors and stores them as a model data TB 52. When determining an article to be recommended, the CPU obtains an attribute vector or a characteristic vector Ui of the target user, which shows state of the target user, and calculates a similarity S<SB>i</SB>to an event probability vector V(P)i of a corresponding factor in the article analysis result TB 511. The CPU multiplies the similarity S<SB>i</SB>by the biases A(P)i of factors in the model data TB 52 to calculate recommendation levels r<SB>i</SB>for each factor. The CPU calculates a recommendation level of the article from recommendation levels r<SB>i</SB>for all factors. <P>COPYRIGHT: (C)2010,JPO&INPIT

Description

  The present invention relates to an information recommendation device that recommends information suitable for a user.

  In recent years, with the development of information technology and the spread of the Internet, information sources have been diversified and a large amount of information has been accumulated. For this reason, information filtering techniques for organizing a large amount of accumulated information and appropriately obtaining information desired by a user have been studied. For example, in addition to a method of statistically organizing information (statistical method), a method of evaluating the appearance frequency of keywords included in document information for the entire document including other documents (content-based filtering; Content-based filtering, a technique for evaluating similarity between a user and another person and extracting appropriate information (collaborative filtering; collaborative filtering) and the like have been developed. Also, a recommendation technique (information recommendation) for recommending information suitable for a user using the above-described information filtering technique has been proposed.

For example, in the above-described method based on collaborative filtering, when recommending information about a certain product to a user, the product purchase history information collected at a POS (Point Of Sales) or EC (Electronic Commerce) site is used. A purchasing user is specified, and a product to be recommended is determined by paying attention to the similarity between the purchasing user and the recommendation target user (Non-Patent Document 1).
Further, in the information recommendation server device of Patent Document 1, when recommending a document to a recommendation target user, the content of the document is expressed by a document vector by the above-described method based on the content-based filtering and accessed by each user. A user profile is created based on a document vector of a document, and a collaborative user similar to the target user is identified from the above user profile by a method based on collaborative filtering, thereby recommending information suitable for the user's preference (Patent Literature) 1).

"Web mining and information theory learning theory" Kenji Yamanishi IPSJ "2002 Informatics Symposium" Online Proceedings " JP 2002-215665 A

On the other hand, when a user's action result is arranged according to various factors affecting the action result, the bias of the factor may be different for each action result.
For example, taking the behavioral result of purchasing a product as an example, some products may be biased in factors such as the sex, age, and region of the purchaser. For example, cosmetics buyers are biased towards women (ie gender factors). In addition, for example, when purchasing content such as books, CDs, and DVDs, there is a bias in interest factors such as the author and category. In addition, products that are easy to buy together with certain products, such as game machines and game software, personal computers and peripheral devices, have a bias in the cause of simultaneous purchase.
There has been a problem that if a product to be recommended is determined without paying attention to such a bias of factors, it is not always an appropriate recommendation.

  The present invention has been made in view of such problems, and the object of the present invention is various factors such as attributes and interests of the target user, and various factors attributable to the product (product attribute information and association). In view of the bias of (interesting information), it is to provide an information recommendation device that recommends appropriate information to the user.

  In order to solve the above-described problem, a first invention is an information recommendation device for recommending information to a target user, and information on users including the target user and other users from a predetermined information source, and A means for acquiring information related to the product and storing the user's action results performed on the product in association with a plurality of factors caused by the user or the product, and an event probability vector for each factor related to the product of interest An event probability calculating means to be calculated; a factor evaluating means for determining the degree of bias of each factor relating to the product of interest; a user characteristic calculating means for determining an attribute vector or a characteristic vector for each factor of the target user; and Based on degree, event probability vector for each factor, and attribute vector or characteristic vector for each factor of the target user An information recommendation apparatus comprising: the recommendation evaluation means for determining whether to recommend products that the focusing on the subject user.

  In addition, the recommendation evaluation unit evaluates the similarity between the event probability vector for the factor having a relatively large degree of bias and the attribute vector or characteristic vector of the factor of the target user, thereby focusing on the product of interest It is desirable to determine whether or not to be a recommended product.

  In addition, it is desirable to further include an information recommendation unit that presents the recommended product determined by the recommendation evaluation unit together with a factor that serves as a basis for recommendation.

  A second invention is an information recommendation method for recommending information to a target user, and acquires information about a user including the target user and another user, and information about a product from a predetermined information source, An accumulation step of accumulating the user's action result performed on the product in association with a plurality of factors caused by the user or the product, an event probability calculating step for obtaining an event probability vector for each factor related to the product of interest, Factor evaluation step for obtaining the degree of bias of each factor related to the product of interest, user characteristic calculation step for obtaining an attribute vector or characteristic vector for each factor of the target user, the degree of bias of each factor, Based on the event probability vector and the attribute vector or characteristic vector of each factor of the target user, the target user. A recommendation evaluation step of determining whether to recommend products that the interest in The, information recommendation method, which comprises a.

  A third invention is a program for causing a computer to function as the information recommendation device of the first invention.

  According to the present invention, it is possible to provide an information recommendation device that recommends appropriate information to a target user in view of various factors such as user attributes and interests and bias of various factors caused by a recommendation target.

Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the drawings.
FIG. 1 is a diagram showing a system configuration of an information recommendation system 1 including an information recommendation device according to the present invention.

  As shown in FIG. 1, the information recommendation system 1 includes a target user PC 2, a master data DB (database) 3, a POS server 4, an information recommendation device 5, a WEB server 6, and the like. Communication connection is established.

  The master data DB 3 includes attribute information (hereinafter referred to as user information) regarding users including target users and other users, attribute information regarding products (hereinafter referred to as product information), attribute information regarding stores (hereinafter referred to as store information), and the like. The database to store.

The POS server 4 is a server device connected to a cash register terminal or the like of an actual store, and collects purchase histories collected at each cash register terminal as purchase history log data 41.
The WEB server 6 is a server device such as an EC site that is open to the WEB, and collects an access history to the EC site, a purchase history of products, a browsing history, and the like as purchase history log data 61.

It is assumed that at least one or a plurality of POS servers 4 and WEB servers 6 exist in the information recommendation system 1.
In the information recommendation system 1, the information recommendation device 5 recommends appropriate information to the target user PC2 in response to a request from the target user PC2.
In the present embodiment, recommended information is mainly product information. The product is not limited to goods, and includes travel products, financial products, and the like. The recommended information (product) may include information such as location information (medical institution, store, etc.), content data (homepage, blog), or services.

  As shown in FIG. 1, the information recommendation apparatus 5 includes a purchase analysis result DB 51, model data TB 52, user profile 53, storage means 54, event probability calculation means 55, factor evaluation means 56, user characteristic calculation means 57, and information recommendation. Means 58 are provided.

  In the following description, “database” may be abbreviated as “DB”, “table” as “TB”, and “profile” as “PF”. “Keyword” is abbreviated as “KW”.

  The storage means 54 of the information recommendation device 5 stores user information, product information, store information, and the like stored in the purchase history log data 41 and 61 collected by the POS server 4 and the WEB server 6 and the master data DB 3. Obtain information from various sources. Specifically, the storage unit 54 creates various profiles (see FIG. 13) from the purchase history log data 41 and 61 of each user and stores them in the user profile 53. Further, the storage means 54 determines the user's action result based on the product purchase history log data 41 and 61, the user information 31, etc. (for example, gender, age, residential area, interest KW information, interest category information). , A product analysis result TB511 (see FIG. 7) associated with the product purchased together). Here, the user's action result means purchase or browsing of a product. In the following description, an event that appears as each factor is referred to as a factor event. For example, in the factor “gender”, the factor event is {male, female, unknown}.

  The event probability calculation means 55 obtains an event probability vector V (P) i for each factor i related to the product P of interest based on the information accumulated by the accumulation means 54. Here, the event probability vector V (P) i is expressed as a set of the ratios of the factor events of each factor (hereinafter referred to as factor values). The factor value calculated by the event probability calculating means 55 is stored in the product analysis result TB511.

  The factor evaluation means 56 calculates the degree of bias A (P) i of each factor i related to the product P of interest and stores it in the model data TB52. Details of the calculation of the factor value and the degree of bias will be described later.

The user characteristic calculation means 57 calculates the attribute vector or characteristic vector Ui of each factor of the target user. The attribute vector or characteristic vector Ui of each factor of the target user is, for example, for the factor “gender”, when the target user is male, U _sex = {1,0,0}, and when female, the U _sex = 0, 1, 0} The attribute value uniquely determined by the user attribute is arranged in the form of the event probability vector V (P) i of the factor. Similarly, for the factors “age” and “region”, the attribute value of the user is fixed, and the number of possible factor events is determined, so that the attribute vector U _age and U _{residence area} can be determined respectively. In addition, for the factors “interest KW information”, “interest category information”, and “joint purchase information”, the tendency of the target user is aligned in the form of the event probability vector V (P) i, and vector data The vector U _{interest KW} , U _{interest category} , and U _{buy information} . In the following description, the attribute vector and the characteristic vector have the same function as vectors representing the user's tendency, and therefore, both are described as Ui.

The recommendation evaluation unit 58 includes the event probability vector V (P) i calculated by the event probability calculation unit 55, the degree of bias A (P) i of each factor i calculated by the factor evaluation unit 56, and the user characteristic calculation unit. Based on the attribute vector or characteristic vector Ui of each factor i of the target user calculated by 57, it is determined whether or not to recommend the product P focused on the target user.
When determining whether or not to recommend, the recommendation evaluation means 58 uses the event probability vector V (P) i for the factor i related to the product P, the target i for the relatively large degree of bias A (P) i, and the target The similarity S with the attribute vector or characteristic vector Ui for the user factor i is calculated. Further, by calculating the recommendation degree R from the similarity S and the degree of bias A (P) i, it is determined whether or not the product P is to be a recommended product.

  The information recommendation device 5 presents the recommended product determined by the recommendation evaluation unit 58 to the target user. That is, among the factors related to the purchase of merchandise, the similarity is more severely evaluated and the similarity is reflected in the recommendation information. Further, when presenting a recommended product, the information recommendation device 5 also presents a basis for recommending the product.

Next, the hardware configuration of the information recommendation device 5 will be described with reference to FIG.
As shown in FIG. 2, the information recommendation device 5 includes, for example, a control unit 501, a storage unit 502, a media input / output unit 503, an input unit 504, a printing unit 505, a display unit 506, a communication unit 507, and the like via a bus 508. Connected and configured.

The control unit 501 includes a CPU (Central Processing Unit), a ROM (Read Only Memory), a RAM (Random Access Memory), and the like.
The CPU calls a program stored in the storage unit 502, ROM, recording medium, or the like to a work memory area on the RAM and executes it, and drives and controls each unit connected via the bus 508. The CPU of the control unit 501 executes model data creation processing (see FIG. 17) and recommended product determination processing (see FIG. 19) described later. The CPU 501 functions as the storage unit 54, event probability calculation unit 55, factor evaluation unit 56, user characteristic calculation unit 57, and recommendation evaluation unit 58 of FIG.

  The ROM permanently holds a computer boot program, a program such as BIOS, data, and the like. The RAM temporarily stores the loaded program and data, and includes a work area used by the control unit 501 for performing various processes.

The storage unit 502 is an HDD (hard disk drive), and stores a program executed by the control unit 501, data necessary for program execution, an OS (operating system), and the like. These program codes are read by the control unit 501 as necessary, transferred to the RAM, and read and executed by the CPU.
The storage unit 502 stores a purchase analysis result TB51, model data TB52, user profile 53, and the like generated by the CPU.

  The media input / output unit 503 (drive device) is a media input / output device such as a floppy (registered trademark) disk drive, PD drive, CD drive, DVD drive, and MO drive, and performs data input / output.

  The input unit 504 is, for example, a keyboard, a pointing device such as a mouse, or an input device such as a numeric keypad, and outputs input data to the control unit 501.

A printing unit 505 is a printer, and performs printing processing such as a recommendation report 11 (see FIG. 25).
The display unit 506 includes a display device such as a liquid crystal panel or a CRT monitor, and a logic circuit (such as a video adapter) for executing display processing in cooperation with the display device, and is input under the control of the control unit 501. Display information is displayed on a display device.
The communication unit 507 includes a communication control device, a communication port, and the like, and is a communication interface that mediates communication with the network 9 and performs communication control.

  Next, the configuration of each data used or generated in the information recommendation system 1 shown in FIG. 1 will be described with reference to FIGS.

As shown in FIG. 3, the master data DB 3 includes user information TB31, product information TB32, store information TB33, and the like.
In the user information TB31, attribute information of each user is stored. That is, data such as user ID 311, gender 312, age 313, and living area 314 of one user per record is stored.

  Specifically, as shown in the user information TB31 in FIG. 4, for example, the record of the user ID 311 “u0001” is associated with the gender 312 “male”, the age 313 “20”, and the living area 314 “Tokyo”. In the record of the user ID 311 “u0002”, the gender 312 “male”, the age 313 “30”, and the living area 314 “Hokkaido” are stored in association with each other.

In the product information TB32 of the master data DB 3 in FIG. 3, attribute information of each product is stored. That is, data such as a product ID 321, a product name 322, a price 323, a manufacturer 324, a description 325, and a product category 326 is stored for one product per record.
Store information TB33 stores attribute information of each store. That is, data such as store ID 331 and area 332 of one store per record is stored.

  FIG. 5 is a diagram showing a data structure of the purchase history TB 43 included in the purchase history log data 41 of the POS server 4 or the purchase history log data 61 of the WEB server 6.

  The purchase history TB 43 stores data such as a cart ID 431, a date and time 432, a product ID 433, a quantity 434, a total 435, a payment method 436, a store ID 437, and a user ID 438 for each purchase (shopping cart) per record.

  Specifically, as shown in the purchase history TB43 of FIG. 6, the record of the cart ID 431 “c10001” includes the date and time 432 “8/10”, the product ID 433 “p002”, the quantity 434 “2”, and the total 435 “10000”. ”, Payment method 436“ cash ”, store ID 437“ s01 ”, and user ID 438“ u0001 ”are stored in association with each other, and the record of the cart ID 431“ c10002 ”includes the date and time 432“ 8/10 ”and the product ID 433“ p003 ”. , Quantity 434 “2”, total 435 “10000”, payment method 436 “cash”, store ID 437 “s01”, and user ID 438 “u0002” are stored in association with each other.

  FIG. 7 is a diagram illustrating a data configuration of the purchase analysis result DB 51 of the information recommendation device 5. As shown in FIG. 7, the product analysis result TB511, product interest KW information TB512, and product interest category information TB513 are stored in the purchase analysis result DB51.

  The product analysis result TB511 is a data group generated in a model data creation process (FIG. 17) described later, which is executed by the CPU of the information recommendation device 5, and includes purchase history log data 41 and 61, product interest KW information TB512, In addition, based on the product interest category information TB513, the purchased factor of each product is accumulated in association with the factor event and the factor value (event probability).

  That is, the product analysis result TB511 includes a plurality of factors 5112 (for example, gender, age, residence area, interest KW information, interest category information, concurrent purchase information) caused by the user or the product regarding the purchase of each product ID 5111, A factor event 5113 that is an event that can be taken by the factor 5112 and a factor value 5114 (an event probability indicating the ratio of the factor event) are stored in association with each other.

FIG. 8 shows a product analysis result TB511 regarding the purchase of the product with the product ID “p010”. Focusing on the first record to the second record of the product analysis result TB511, the analysis result on the factor 5112 “gender” can be referred to. That is, among the users who have purchased the product “p010”, the event probability (factor value) of “male” is “0.56”, and the event probability (factor value) of “woman” is “0. 44 ".
Similarly, from the 3rd to 6th records, the age group of the purchasing user and the event probability (factor value) are known, and from the 7th to 11th records, the living area of the purchasing user and the event probability thereof. (Factor value) is known.

  Focusing on the factor 5112 “interest KW information” in the 12th to 19th records, the cause events are “Star Wars”, “Harry Potter”, “Record W”, “Blu-ray”, “Hi-Vision”, “SD” “Card”, “HDD”, “500 GB”, and the like are stored, and their event probabilities (factor values) are “0.08”, “0.08”, “0.25”, “0. 17 ”,“ 0.17 ”,“ 0.08 ”,“ 0.08 ”, and“ 0.08 ”are stored. These mean the interest keyword associated with the product “p010” and its event probability.

Similarly, from the 20th to 25th records, the interest category associated with the product “p010” and its factor value are known, and from the 26th to 29th records, the concurrent purchase associated with the product “p010”. You can see the product and its event probability.
The co-purchased product related to the product “p010” is a product purchased at the same time as p010.

  Here, in this embodiment, the factor 5112 associated with a certain product is “sex”, “age”, “residential area”, “interest KW information”, “interest category information”, “combined purchase product”. Although shown, it is not limited to this, and other factors may be added or reduced as necessary.

  In addition, as for factors such as “gender”, “age”, “region” and the like, possible events (factor events) are limited, but “interest KW information”, “interest category information”, “jointly purchased products” For such factors, possible events (cause events) are not limited. In that case, the number may be limited and extracted in advance, for example, the top 100 at the time of data collection.

  In the product interest KW information TB 512 shown in FIG. 9, the cart ID 5121, the product ID 5122, and the interest KW information 5123 are stored in association with each other. The interest KW information is a keyword or the like associated with the product, for example, a search keyword or the like. The merchandise interest KW information TB 512 is collected from the user's access history and purchase history (included in the purchase history log data 61).

  Specifically, as shown in the product interest KW information TB 512 in FIG. 9, the cart ID 5121 “c10021” includes product ID 5122 “p010”, interest KW information 5123 “Harry Potter”, “W Record”, “Blu-ray”. Associated.

  In the product interest category information TB513 shown in FIG. 10, the cart ID 5131, the product ID 5132, and the interest category information 5133 are stored in association with each other. Interest category information is a category associated with a product. The merchandise interest category information TB513 is collected from the user's access history and purchase history (included in the purchase history log data 61).

  Specifically, as shown in the product interest category information TB513 of FIG. 10, the cart ID 5131 “c10021” is associated with the product ID 5132 “p010”, the interest category information 5133 “AV device”, and “PC”.

  FIG. 11 is a diagram showing a data configuration of the model data TB52 of the information recommendation device 5. As shown in FIG. The model data TB52 shown in FIG. 11 is a data group generated in a model data creation process (see FIG. 17), which will be described later, executed by the CPU of the information recommendation device 5, and includes purchase history log data 41 and 61, product interests. Based on the KW information TB 512 and the product interest category information TB 513, the degree A (P) i of the bias of the factor i of each product and the representative factor flag are determined. That is, the model data TB 513 stores a product ID 521, a factor 522, a factor bias degree 523, a representative factor flag 524, and the like in association with each other.

  Here, the degree of factor bias (A (P) i) 523 is a numerical value indicating the variation of the factor event, and is based on a standard deviation of the factor value (event probability) or a value inversely proportional to the number of factor events. Calculated.

  Specifically, for example, when the factor is “gender”, the factor event is {male, woman, unknown}, and when the frequency of the factor event is {5, 4, 0}, the factor value of each factor event Is {5/9, 4/9, 0}. Here, the factor value 0 is not included in the calculation. Since the average of the factor values (x with a bar in equation (1)) is 0.5, the degree of factor bias 523 is obtained as “0.66” from the standard deviation of the factor values (equation (1)). reference).

  Further, for example, for the factor “interest KW”, in the example shown in FIG. 9, the factor events are eight items of {Star Wars, Harry Potter, W Record, Blu-ray, Hi-Vision, SD card, HDD, 500 GB}. The frequency of each factor event is {1, 1, 3, 2, 2, 1, 1, 1}. The factor values are {1/12, 1/12, 3/12, 2/12, 2/12, 1/12, 1/12, 1/12}, and the degree of bias of the factor (A (P) i) 523 “0.06” is obtained from the standard deviation of the factor values. The degree of factor bias (A (P) i) 523 may be a value proportional to the reciprocal (1/8) of the number of factor events.

In the model data TB52 shown in FIG. 12, for example, focusing on the first to sixth records, as factors 522 of the product ID 521 “p010”, “sex”, “age”, “residential area”, “interesting KW information” , “Interest category” and “Buy purchased products”, the degree of factor bias (A (P) i) 523 is “0.06”, “0.24”, “0.04”, “0.06”, respectively. ”,“ 0.07 ”,“ 0.15 ”and the like are stored. Further, the representative factor flag 524 of “age” (second record) having the largest factor bias degree (A (P) i) 523 is set to “1”. The representative factor flag 524 of the other record of the product “p010” is set to “0”.
This means that the purchase of the product “p010” is most dependent on the “age” factor having the largest degree of bias A (P) i523.

FIG. 13 is a diagram illustrating data included in the user profile 53 of the information recommendation device 5. The user profile 53 is a data group accumulated by the CPU of the information recommendation device 5 based on the user's access history to the WEB server 6, product browsing / purchasing, and the like.
In other words, the user profile 53 stores user interest KW information PF531, user interest category information PF532, user browsing information PF533, and the like.

As shown in FIG. 14, in the user interest KW information PF 531, a user ID 5311, interest KW information 5312, the number of purchases / viewings 5313, and a score 5314 are stored in each record. Here, the score 5314 is a ratio (event probability) of the number of purchases / views 5313. For example, the total number of purchases / views of the user with the user ID 5311 “u0100” is 32 times. The number of purchases / views in the interest KW “Record W” is “10”, and the score 5314 is calculated as “10/32 = 0.31”.
The user interest KW information PF 531 is updated every time the user accesses.

As shown in FIG. 15, in the user interest category information PF 532, a user ID 5321, interest category information 5322, the number of purchases / viewings 5323, and a score 5324 are stored in each record. Here, the score 5324 is a ratio (event probability) of the number of purchases / views 5323. For example, the user ID 5321 “u0100” is purchased and viewed a total of 38 times. The number of purchases / viewings in the interest category “AV device” is “12”, and the score 5324 is calculated as “12/38 = 0.32.”
The user interest category information PF 532 is updated every time the user accesses.

As shown in FIG. 16, in the user browsing information PF533, a user ID 5331, a product ID 5332, the number of purchases / viewings 5333, and a score 5334 are stored in each record. Here, the score 5334 is a ratio (event probability) of the number of purchases / views 5333. For example, the user ID 5331 “u0100” has 8 purchases / views in total. The number of purchases / views 5333 of the product “p001” is “2”, and the score 5334 is calculated as “2/8 = 0.25”.
The user browsing information PF 533 is updated every time a user accesses.

  Next, the flow of model data creation processing and recommended product determination processing executed by the information recommendation device 5 will be described with reference to FIGS.

  First, at the stage of executing model data creation processing and recommended product determination processing, user information TB31, product information TB32, and store information TB33 are collected in the master data DB 3, and the POS server 4 or WEB server 6 stores FIG. It is assumed that the purchase history TB43 is stored. Further, it is assumed that user interest KW information 531 in FIG. 14, user interest category information 532 in FIG. 15, and user browsing information 533 in FIG. 16 are generated and stored in the user profile DB 53 of the information recommendation device 5. Further, it is assumed that the merchandise interest KW information 512 of FIG. 8 and the merchandise interest category information 513 of FIG. 9 are stored in the purchase analysis result DB 51 of the information recommendation device 5.

  In the model data creation processing of FIG. 17, the CPU of the control unit 501 of the information recommendation device 501 acquires the purchase history TB 43 (FIG. 6) from the purchase history log data 41 of the POS server 4 or the purchase history log data 61 of the WEB server 6. Then, a list of purchased products is generated (step S101).

  The concurrent purchase product list 7 shown in FIG. 18 is obtained by extracting the concurrent purchase product of the product p010 from the purchase history TB43 in FIG. As shown in FIG. 18, the commodities purchased together include a product ID {p001, p002, p003, p004}, and the purchased user is a user ID {u0001, u0002, u0003, u0004, u0005, u0006, u0007, u0008, u0009. }.

  Next, the CPU 101 repeatedly executes the following processing (step S103 to step S107) in order from the product ID “p001” (step S102).

  In step S103, the CPU associates the product with the factors (age, gender, living area) included in the user information TB31 (see FIG. 4), calculates the factor value (event probability) of each factor event, and the product shown in FIG. Stored in the analysis result TB511.

That is, first, the CPU 101 refers to a record related to the user IDs {u0001, u0002, u0003, u0004, u0005, u0006, u0007, u0008, u0009} of the user information TB31 in FIG. 4 to determine “age”, “sex”, “ Acquire the cause event (including the cause event and the frequency of each cause event) of the “residential area”. For “gender”, the cause event is {male, female, unknown}, and the frequency is {5, 4, 0}.
Next, the CPU calculates a factor value (event probability) from the frequency of the factor event. The factor value is the ratio of each factor event, and {5/9 = 0.56, 4/9 = 0.44, 0} is obtained. Further, the degree A (P) i of the factor bias is obtained using the above-described equation (1).

Similarly, for the factor “age”, the factor event is {not in teens, teens, 20s, 30s, 40s, 50s, over 60s}, and the frequency is {0, 0, 1 , 6, 1, 1, 0}.
The factor values are {0, 0, 1/9 = 0.11, 6/9, = 0.67, 1/9 = 0.11, 1/9 = 0.11, 0}. Note that FIG. 8 does not describe data in which the frequency of the cause event is “0”.

Similarly, for the factor “residential area”, the factor event is {Tokyo, Hokkaido, Kanagawa, Saitama, Chiba}, and the frequency is {2, 2, 2, 2, 1}.
The factor value is {2/9 = 0.22, 2/9 = 0.22, 2/9 = 0.22, 2/9 = 0.22, 1/9 = 0.11}.

  Next, in step S104, the CPU associates the product with the interest KW {Star Wars, Harry Potter, W Record, Blu-ray, Hi-Vision, SD card, HDD, 500GB} included in the product interest KW information 512 of FIG. The factor value of the factor is calculated and stored in the product analysis result TB511 in FIG.

That is, the CPU refers to the record related to the product p010 in the product interest KW information TB512 shown in FIG. And determine the frequency of the causal event. The frequency is {1, 1, 3, 2, 2, 1, 1, 1}.
Next, the CPU calculates a factor value (event probability) from the frequency. Each factor value (event probability) of interest KW information is {1/12 = 0.08, 1/12 = 0.08, 3/12 = 0.25, 2/12 = 0.17, 2/12 = 0. .17, 1/12 = 0.12, 1/12 = 0.08, 1/12 = 0.08}.

  Next, in step S105, the CPU associates the product with the product interest category included in the product interest category information 513 shown in FIG. 10, calculates the factor value of each factor, and stores it in the product analysis result TB511 of FIG.

That is, first, the CPU refers to the record regarding the product p010 in the product interest category information TB513 shown in FIG. 10, and the frequency of each factor event of the product interest category {AV device, DVD, consumer electronics, personal computer, game, clock}. Ask for. The frequency is {3, 1, 2, 3, 1, 1} with reference to FIG.
Next, the CPU calculates a factor value from the frequency. Each factor value (event probability) of the interest category information is {3/11 = 0.27, 1/11 = 0.09, 2/11 = 0.18, 3/11 = 0.27, 1/11 = 0. .09, 1/11 = 0.09}.

  Next, in step S106, the CPU calculates a factor value (event probability) of each factor based on the extracted concurrent purchase product list 7 as a factor, and stores it in the product analysis result TB511 of FIG.

  That is, the CPU obtains each frequency {2, 2, 5, 1} of the co-purchased product {p001, p002, p003, p004} from the list 7 of the co-purchased product in FIG. The factor values (event probabilities) are {2/10 = 0.2, 2/10 = 0.2, 5/10 = 0.2, 1/10 = 0.1}.

  Next, in step S107, the CPU calculates the degree of bias A (P) i of each factor from each factor value calculated in the processing in steps S103 to S106, and stores the model data TB513 as shown in FIG. Generate. As the degree of factor bias A (P) i, the standard deviation shown in the above equation (1) may be used, or a value inversely proportional to the number of factor events may be used.

  Further, the CPU sets a representative factor flag 5134 for each product. In the present embodiment, for the representative factor flag 524, for example, the factor having the largest factor deviation degree A (P) i is set to “1”, and the other factors are set to “0”. For the product “p010”, the representative factor flag 5134 of the factor “age” is set to “1”.

In FIG. Model data TB52 illustrated in 12, deflection degree A (P) _sex factors product "p010", "sex" is "0.06", the degree A (P) _age bias factors "age" is "0.24", factors the degree of a (P) _{residential area} of the bias of the "living area" is "0.04", factors the degree of a (P) _{interest KW} of bias in the "interest KW information" is "0.06", factors the degree of a (P) _{interest category} of bias in the "interest category information" is "0.07", factor "併買products" degree of a (P) _{併買commodity} bias of " 0.15 ". In the representative factor flag of the product “p010”, the factor “age” is set to “1”.

  As described above, the factor value (event probability) of the factor and the degree A (P) i of the factor bias are calculated for all the products, and the product analysis result TB511 and the model data TB52 are generated (step S102; Yes). ), The CPU stores the generated product analysis result TB511 and model data TB52 in the storage unit 402, and ends the series of model data creation processing.

  Next, the information recommendation device 5 executes processing for determining a product recommended for the target user based on the generated product analysis result TB511 and the model data TB52.

FIG. 19 is a flowchart showing the flow of recommended product determination processing.
As shown in FIG. 19, first, the CPU extracts the user characteristic 8 of the target user and obtains a user attribute vector or characteristic vector Ui (step S201). As shown in FIG. 20, the user characteristic is a score given to each factor associated with the target user (here, the user with the user ID “u0100”). The score is a numerical value indicating the strength of the association between the user and the factor event, and is given in the range of 0 to 1, for example. When the user is related to only one factor event (for example, when the factor is “gender”), the score is “1”. When the user has a plurality of factor events (for example, in the case of factor “interest KW information”), the event probability becomes a score. The CPU aligns the score of each factor calculated for the target user in the form of the event probability vector V (P) i of each factor for the product, and sets it as the attribute vector or characteristic vector Ui of the target user.

  The first to third records of the user characteristic 8 shown in FIG. 20 are generated and stored based on the user information TB31 (FIG. 4). The fourth to sixth records are stored based on the user interest KW information PF531 (FIG. 14). The seventh to ninth records are stored based on the user interest category information PF532 (FIG. 15). The 10th to 15th records are stored based on the user browsing information PF533 (FIG. 16).

  Scores 5314 and 5324 of the user interest KW information PF 531 and the user interest category information PF 532 are stored in the scores of the factors “interest KW information” and “interest category information” of the user characteristic 8. Note that it is not necessary to extract all the factor events stored in the user interest KW information PF531 and the user interest category information PF532, and it is also possible to extract those whose scores are equal to or greater than a predetermined value (can be arbitrarily set). . In the example of the user characteristic 8 in FIG. 20, those having a score of “0.2” or more are extracted. In addition, for the “viewing product” of the user characteristic 8, all the corresponding products in the user browsing information PF533 are extracted regardless of the score.

Next, the CPU calculates a recommendation degree R for all products (excluding products viewed by the target user) in order from the product ID “p001” (step S202).
As a specific example, calculation of the recommendation level R for the product “p010” that is not included in the browsed products of the target user “u0100” will be described.

The degree of recommendation R is the degree of similarity between the attribute vector or characteristic vector Ui of each factor of the target user and the event probability vector V (P) i of each factor i of the product analysis result TB511, and the degree of bias A (P ) I is calculated in consideration of i. For example, the sum of those multiplied by the deviation A (P) i factors in the similarity _{S i} of each factor _{(S i * A (P)} i = r i) or as a recommended degree R, representative factor flag Only the similarity S _{i of the} factor set to “1” may be obtained and used as the recommendation R.

First, the CPU calculates a similarity S _i .
The similarity S _i is calculated as follows from the event probability vector V (P) i of each factor (for example, gender) associated with the product P and the attribute vector or characteristic vector Ui of the target user.

That is, the similarity S _i is expressed by the inner product of the event probability vector V (P) i of the product and the attribute vector or characteristic vector Ui of the target user for a certain factor (for example, gender).
The similarity S _i for the i-th factor is calculated using the following equation (2).

Here, Ui is an attribute vector or characteristic vector for the factor i of the target user,
Ui = (a _ui , b _ui , c _ui ,...).
V (P) i is an event probability vector of the factor i of the product P of interest,
V (P) i = (a _i , b _i , c _i ,...).
_{a ui, b ui, c ui} , ··· corresponds to the score of the user characteristic _{8, a i, b i,} c i, ··· correspond to the cause value of the product analysis results TB511.

After calculating the similarity S _i , the CPU calculates a recommendation R.
First, as a recommended degree r _i factors i, as shown in equation (3), the calculated similarity S _i, multiplying the degree A (P) i bias factors.

r _i = A (P) i * S _i (3)

The recommendation level R of the product may be the recommendation level r _i for the factor for which the representative factor flag is set to “1”, or may be calculated by the equation (3) as expressed by the following equation (4). it may be that the sum of the recommendation degree r _i for all of the factors were.

A specific calculation example of the degree of similarity and the degree of recommendation will be described with reference to FIGS. 21 and 22.
The target product p010 has an event probability vector V (P) i and a factor bias degree A (P) i as shown in the left table of FIG. 21, and the target user has attributes as shown in the right table of FIG. Assume that a vector or characteristic vector Ui is included.
Using FIG. 21 and the above equation (2), the similarity S _i as shown in FIG. 22 is calculated. Also, the recommendation degree R is calculated from the similarity S _i and the degree of bias A (P) i of the factor i.

Specifically, by substituting the numerical values shown in FIG. 21 into the above formulas (2), (3), and (4), values as shown in the table of FIG. 22 are calculated.
For example, the similarity S _i for the factor “gender” is calculated by substituting V (P) i and Ui in FIG. 21 into the above equation (2), as shown in equation (5):

  Thus, “0.79” is calculated.

Further, the recommendation degree r _i is obtained by substituting the degree of bias A (P) i of the factor in FIG.
r _i = 0.06 * 0.79
= 0.047 (6)
It becomes.

Further, for example, the similarity S _i for the factor “interesting KW information” is obtained by substituting V (P) i and Ui in FIG. 21 into the above equation (2), as shown in equation (7):

  Thus, “0.73” is calculated.

Further, the recommendation degree r _i is obtained by substituting the degree of bias A (P) i of the factor in FIG.
r _i = 0.07 * 0.73
= 0.046 (8)
It becomes.

Similarly, the similarity S _i and the recommendation degree r _i are calculated for other factors (“age”, “residential area”, “interesting category information”, “buying information”), and the recommendation degree R is obtained.

  When the CPU finishes calculating the recommendation level for the target product currently being calculated, when there is a target product to be calculated next (step S202; No), the CPU returns to step S203 to determine the next calculation target product. The degree of recommendation is calculated. When the recommendation level for all the target products is calculated (step S202; Yes), the CPU determines the recommended product from the recommendation level and presents it with the reason for recommendation (step S204).

FIG. 23 is an example of the recommended product list 9 determined by the CPU, and products having a high recommendation level are arranged in order along with representative factors.
For example, the recommendation degree of the first recommended product p010 is “0.427”, and the representative factor is “age”. Further, the recommendation degree of the second recommended product p020 is “0.411”, and the representative factor is “interesting KW information”. In addition, the recommendation degree of the third recommended product p008 is “0.383”, and the representative factor is “both purchase information”.

  When presenting the recommended product in step S204, the CPU generates a recommended document 10 (see FIG. 24) and a recommendation report 11 (see FIG. 25) for recommending the recommended product based on the calculated recommended product list 9. To do.

  The recommended document 10 shown in FIG. 24 is described in, for example, a markup language for creating a WEB page such as HTML, XML, etc., and product images 10a, for example, for the top three products in the recommended product list 9. 10b and 10c are posted, and links 10g, 10h, and 10i to the recommendation reason page are provided. The reasons for recommendation 10g, 10h, and 10i may be displayed together with the recommended product images 10a, 10b, and 10c in the recommended document 10 at the same time.

  The recommendation document 10 is provided to the EC site 6 or the like, for example, and is displayed on the top page when the target user PC2 accesses the EC site 6, or the target user PC2 searches for a predetermined product. When entered as a keyword, the recommended document 10 is displayed.

  Further, as in the recommendation report 11 shown in FIG. 25, the relationship between the campaign product X and other products is investigated by the same method (model data creation processing, recommended product determination processing), and products with a high recommendation degree are related products. The reasons 11d, 11e, and 11f for picking up the product may be described. Since the recommendation report 11 is a report related to “store display”, unlike the recommended document 10 in FIG. 24, the target user is not specified. In this case, instead of the attribute factors (gender, age, residence area) of the target user, the factors related to the attributes of the campaign product X (product name, price, manufacturer, product category, etc .; see FIG. 3) Data TB52 and product analysis result TB511 are created, and the recommendation level of each product may be calculated based on the similarity and the bias of factors.

As described above, in the information recommendation system 1, the CPU of the information recommendation device 5 acquires attribute data and purchase history log data 43 and the like regarding the user and the product from the master data DB 3, the store server 4, the WEB server 6, and the like. The product is associated with a plurality of factors caused by the user or the product, and the event probability vector V (P) i of the factor event of each factor is calculated. Further, the CPU calculates the degree of bias A (P) i of each factor and stores it as model data TB52. Further, when determining the recommended product, the CPU obtains the characteristic (attribute vector or characteristic vector Ui) of the target user and calculates the similarity S _i with the event probability vector V (P) i described above. Further, by multiplying the deviation A (P) i factor model data TB52 similarity _{S i,} and calculates the recommendation degree _{r i} of each factor. The CPU obtains a recommendation degree R of the product from the recommendation degree r _i for all factors, for large items of recommendation level R, generates a recommendation document 10 and recommendation report 11, and outputs.

Therefore, it is possible to recommend appropriate information to the user in view of various factors such as the attributes and interests of the target user and various factors caused by the product.
Moreover, since the recommendation reason is also shown when the recommended product is presented like the recommendation document 10 and the recommendation report 11, the user can easily understand the reason for the recommendation.

  In the above-described embodiment, the factor bias is calculated based on the standard deviation of the factor value or a value inversely proportional to the number of events that can be taken. Any value may be used as long as it exists. For example, “Tokyo” may be set to “0”, “Hokkaido” or “Okinawa” may be set to “1”, and the closeness of the area may be quantified to calculate more detailed bias. .

  In the above-described embodiment, the product analysis result TB51 related to the purchase of the product is generated. However, the present invention is not limited to the purchase, and a product in which some action result of the user performed on the product is associated with each factor. The analysis result TB51 may be used. For example, the browsing (access) behavior of the user with respect to the product may be associated with each factor.

  In addition, it is obvious that those skilled in the art can come up with various changes and modifications within the scope of the technical idea disclosed in the present application, and these naturally belong to the technical scope of the present invention. It is understood.

System configuration example of information recommendation system 1 Hardware configuration example of the information recommendation device 5 Data configuration example of master data DB3 Data example of user information TB31 Data structure example of purchase history log data 41 and 61 Data example of purchase history TB43 Data structure example of purchase analysis result DB51 Data example of product analysis result TB511 Data example of product interest KW information TB512 Data example of product interest category information TB513 Data structure example of model data TB52 Specific example of model data TB52 Data configuration example of user profile 53 Data example of user interest KW information PF531 Data example of user interest category information 532 Data example of user browsing information PF533 Flowchart explaining the flow of model data creation processing An example of a list 7 of simultaneously purchased products extracted for the product p010 Flowchart explaining the flow of recommended product determination processing An example of user characteristics 8 for the target user Table showing each factor, factor event, frequency, event probability vector V (P) i, bias A (P) i for the product p010, and table showing attribute vector or characteristic vector Ui for the target user Table showing similarity and recommendation for product p010 of target user An example of the calculated recommended product list 9 An example of a recommended product recommendation document 10 determined by the information recommendation device 5 An example of a recommended product recommendation report 11 determined by the information recommendation device 5

Explanation of symbols

1 ... Information recommendation system 2 ... Target user PC
3 ... Master data DB
4 ... POS server 5 ... Information recommendation device 6 ... WEB server 511 ... Product analysis result TB
52 ... Model data TB
S _i ... Similarity for factor i V (P) i... Event probability vector of factor i for product P Ui... Attribute vector or characteristic vector of factor i for target user A (P) i ... Degree of bias of factor i r _i ... Degree of recommendation of factor i R ... Degree of recommendation of product

Claims (5)

An information recommendation device for recommending information to a target user,
A plurality of factors resulting from the user or the product from the information about the user including the target user and other users and the information about the product from a predetermined information source, and the user's action result performed on the product. Storage means for storing in association with
An event probability calculating means for obtaining an event probability vector for each factor relating to the product of interest;
Factor evaluation means for determining the degree of bias of each factor related to the product of interest,
User characteristic calculation means for obtaining an attribute vector or characteristic vector for each factor of the target user;
Based on the degree of bias of each factor, the event probability vector for each factor, and the attribute vector or characteristic vector of each factor of the target user, it is determined whether to recommend the product of interest to the target user. Recommendation evaluation means,
An information recommendation device comprising:   The recommendation evaluation means recommends a product of interest by evaluating the similarity between the event probability vector for a factor having a relatively high degree of bias and the attribute vector or characteristic vector of the factor of the target user. The information recommendation device according to claim 1, wherein it is determined whether or not the product is a product.   The information recommendation device according to claim 1, further comprising information recommendation means for presenting the recommended product determined by the recommendation evaluation means together with a factor as a basis for recommendation. An information recommendation method for recommending information to a target user,
A plurality of factors resulting from the user or the product from the information about the user including the target user and other users and the information about the product from a predetermined information source, and the user's action result performed on the product. An accumulation step for accumulating in association with
An event probability calculation step for obtaining an event probability vector for each factor related to the product of interest;
A factor evaluation step to determine the degree of bias of each factor related to the product of interest;
A user characteristic calculation step for obtaining an attribute vector or characteristic vector for each factor of the target user;
Based on the degree of bias of each factor, the event probability vector for each factor, and the attribute vector or characteristic vector of each factor of the target user, it is determined whether to recommend the product of interest to the target user. A recommendation evaluation step;
An information recommendation method comprising:   A program for causing a computer to function as the information recommendation device according to claim 1.

Images