JP2014232346A - Information recommendation device, information recommendation method, and information recommendation program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To recommend items suitable for a user's need.SOLUTION: If there are items that fall under a specific-period purchase pattern considering a merchandise purchase period or a regular-interval purchase pattern considering an interval between purchases, identification information of each of the items and identification information of each user in association with each other are connected by an edge. Then, weights according to amounts of elapsed time from respective purchase times are given to respective edges. If there is an item that falls under a predetermined purchase pattern, a graph structure is generated with a weight given to an edge related to the item adjusted, and an item to be recommended is determined based on weights given to the edges in the generated graph structure.

Description

  The present invention relates to a technique for recommending items.

  Conventionally, the usage history of service content items (hereinafter referred to as items) used in the past is accumulated, and the relationship between user information and item information included in the history is regarded as a graph structure. There is a technique for recommending items that the user is interested in.

  In Non-Patent Document 1, user information and item information in the usage history are used as nodes, node connection relationships are represented in a graph structure, and an RWR (Random Walk with Restart) algorithm is applied to identify recommended items. ing.

  In the RWR algorithm, popular items that are highly related to other nodes on the graph are regarded as important items, so that it is possible to recommend both the degree of personalization and the degree of popularity. Even if you are not using enough, you can recommend popular items and use them in a wider range of situations.

  Moreover, in the said nonpatent literature 1, an item is effectively recommended even when there is little utilization history by further considering the metadata information which showed the relationship between users and between items other than user information and item information. Like to do.

  Furthermore, in Non-Patent Documents 2 and 3, focusing on the fact that the most recent usage history is more related to the user's recent interest than the past usage history, using the forgetting function, the progress from the item usage time The weight for interest in the item is attenuated according to time.

I. Kontas, 2 others, “On Social Networks and Collaborative Recommendation”, Proc. SIGIR '09, p.195-202 Y.Ding, 1 other, “Time Weight Collaborative Filtering”, Proc. CIKM ‘05, p.485-492 Nathan N. Liu, 3 others, “Online Evolutionary Collaborative Filtering”, Proc. RecSys ‘10, p.95-102

  However, according to the technique of Non-Patent Document 1, since the most recent usage history and the past usage history are treated equally, when the user's interest changes with the passage of time, the item recommendation accuracy decreases. .

  On the other hand, according to the techniques of Non-Patent Documents 2 and 3, since the degree of interest in the used item is attenuated with the passage of time, it is possible to recommend an item that matches the most recent usage history.

  However, in reality, items used at certain times (such as toys purchased within a few days before the child's birthday) and items used over a certain time span (toothpaste purchased every few months) Etc.), etc., and there is an item for which it is not appropriate to simply attenuate the degree of interest over time, and such an item could not be recommended appropriately.

  The present invention has been made in view of the above circumstances, and an object thereof is to recommend items suitable for the convenience of the user.

  The information recommendation apparatus according to claim 1, a storage unit that stores identification history of an item used by a user and a usage history in which the identification time of the item is associated with the identification information of the user; Identification means for acquiring the usage history of each item and using the usage time included in each usage history to identify a temporal usage pattern in which each item was used for each item or each user, and identification of the associated item Information and the user's identification information are respectively connected by edges, weights corresponding to the elapsed time from each usage time are given to each edge, and if there is an item that matches a predetermined usage pattern, the edge related to that item Generating means for generating a graph structure with corrected weights; and determining means for determining recommended items based on weights given to each edge in the graph structure; And summarized in that with.

  According to the present invention, when there is an item that matches a predetermined usage pattern, the weight of the edge on the graph structure related to the item is corrected, so that an item suitable for the convenience of the user can be recommended.

  The information recommendation device according to claim 2, wherein in the information recommendation device according to claim 1, the generation unit gives each edge a weight that attenuates according to an elapsed time, and the weight of the edge related to the matching item. The gist is to correct the weight to a weight greater than the weight after attenuation.

  According to the present invention, a weight that attenuates according to the elapsed time is given to each edge, and an edge weight related to an item that matches a predetermined usage pattern is corrected to a weight larger than the weight after attenuation. You can recommend items that are more convenient.

  The information recommendation device according to claim 3 is the information recommendation device according to claim 1 or 2, wherein the determination unit uses the graph structure in which metadata about the item is further connected by an edge. The gist is to decide.

  According to the present invention, since a recommended item is determined using a graph structure in which metadata relating to items is further connected by an edge, an item more suitable for the convenience of the user can be recommended.

  The information recommendation method according to claim 4 is a storage step of storing, in a storage unit, a use history in which identification information of an item used by a user and its use time are associated with the identification information of the user by a computer. A specific step of acquiring a plurality of usage histories from the storage means, and specifying a temporal usage pattern in which each item is used for each item or each user using a usage time included in each usage history; When the identification information of the item and the identification information of the user are respectively connected by an edge, a weight corresponding to the elapsed time from each usage time is given to each edge, and there is an item that matches a predetermined usage pattern Includes a generation step for generating a graph structure in which the weight of an edge related to the item is corrected, and a weight given to each edge in the graph structure. The subject matter that has a, a determination step of determining a recommended item based on.

  According to the present invention, when there is an item that matches a predetermined usage pattern, the weight of the edge on the graph structure related to the item is corrected, so that an item suitable for the convenience of the user can be recommended.

  The information recommendation method according to claim 5 is the information recommendation method according to claim 4, wherein the generation step assigns a weight that attenuates according to an elapsed time to each edge, and the weight of the edge related to the matching item. The gist is to correct the weight to a weight greater than the weight after attenuation.

  According to the present invention, a weight that attenuates according to the elapsed time is given to each edge, and an edge weight related to an item that matches a predetermined usage pattern is corrected to a weight larger than the weight after attenuation. You can recommend items that are more convenient.

  The information recommendation method according to claim 6 is the information recommendation method according to claim 4 or 5, wherein the determining step is configured to determine the recommended item using a graph structure in which metadata about the item is further connected by an edge. The gist is to decide.

  According to the present invention, since a recommended item is determined using a graph structure in which metadata relating to items is further connected by an edge, an item more suitable for the convenience of the user can be recommended.

  A gist of an information recommendation program according to a seventh aspect is to cause a computer to execute the information recommendation method according to any one of the fourth to sixth aspects.

  According to the present invention, an item suitable for the convenience of the user can be recommended.

It is a figure which shows the functional block structure of an information recommendation apparatus. It is a figure which shows the example of data of item DB. It is a figure which shows the example of data of service utilization log DB. It is a figure which shows the example of a graph structure. It is a figure which shows the example of data of appearance pattern DB. It is a figure which shows the example of data of recommendation item DB. It is a flowchart which shows an offline process.

  The present invention analyzes the usage pattern of an item from the usage history, and the item corresponding to the predetermined usage pattern is more meaningful to the user by correcting to another weight instead of the weight that attenuates with the elapsed time. There is an item recommendation that does not miss the opportunity to consume items.

  Hereinafter, an embodiment for carrying out the present invention will be described with reference to the drawings. However, the present invention can be implemented in many different modes and should not be construed as being limited to the description of the present embodiment.

  FIG. 1 is a diagram showing a functional block configuration of an information recommendation device 1 according to the present embodiment. The information recommendation device 1 is a device that recommends an electronic or physical item that is likely to be used by a user based on the past usage status of the item of the user.

  Specifically, the browser function unit 11, the content providing unit 12, the item DB 13, the log acquisition unit 14, the service usage log DB 15, the graph construction unit 16, the appearance pattern specifying unit 17, and the appearance pattern DB 18 The first edge weight calculation unit 19, the second edge weight calculation unit 20, the recommendation target specifying unit 21, the recommended item DB 22, and the recommended item presentation unit 23. Hereinafter, the function of each part will be described in detail.

  The content providing unit 12 receives the user ID input by the user via the browser displayed by the browser function unit 11, accesses the item DB 13 based on the internal content providing logic, and acquires the item acquired from the item DB 13 It is a functional unit that presents information to the user via the browser.

  Further, based on the received user ID, the recommended item presenting unit 23 is accessed, and the recommended item acquired from the recommended item presenting unit 23 is presented to the user via the browser, and is determined in advance such as browsing and purchasing. In the case where an item is used in the form as provided, a function of passing the user ID and the item ID of the item used by the user to the log acquisition unit 14 is also provided.

  The item DB 13 is a database that stores item information. An example of item information is shown in FIG. An item ID for identifying an item is stored, and when there is metadata or other data related to each item, the metadata and the like are also stored.

  Note that metadata is additional data representing information about items. For example, when the item is an application program, the name of the program, an explanatory note, a keyword extracted from the explanatory note, and the like. Further, when the item is a proper name of a product, it is a category name such as chocolate.

  The log acquisition unit 14 is a functional unit that receives a combination of a user ID and an item ID from the content providing unit 12 and stores it in the service usage log DB 15 together with time stamp information (use time).

  The service usage log DB 15 is a database that records a user's item usage history. An example of the item usage history is shown in FIG. The combination of the user ID and item ID received by the log acquisition unit 14 is recorded in association with the time stamp information.

  The graph construction unit 16 is based on a pair of user ID and item ID stored in the service usage log DB 15 and has a graph structure in which the user and the item are nodes, and the relationship between the nodes on the service usage log is an edge. Is a function unit that assigns weights specified based on information from the first edge weight calculation unit 19 and the second edge weight calculation unit 20 to each edge and passes them to the recommendation target specifying unit 21.

  Note that the metadata elements of the items stored in the item DB 13 are also handled as nodes according to a predetermined setting. For example, if a keyword that indicates the field of an item is described in metadata, that keyword is also recognized as a node in the graph structure, and a graph structure that is connected to the node indicating the item by an edge is constructed. . An example of the graph structure when the item information in the item DB 13 is shown in FIG. 2 is shown in FIG.

  The appearance pattern specifying unit 17 specifies an item appearance pattern in the entire usage history based on the information in the service usage log DB 15 and the item DB 13, and further specifies an item appearance pattern for each user. It is a functional unit that stores the specific result in the appearance pattern DB 18.

  Examples of the appearance pattern to be specified are a specific period type consumption pattern that does not occur continuously but is consumed at a specific period, a constant span type consumption pattern in which a time span of consumption is assumed to some extent, and the like.

  The former is a case where an item is used at a specific time, such as purchase of chocolate within a few days before Valentine's Day or purchase of a toy within a week before a child's birthday. On the other hand, the latter is a case where items are used at regular intervals, such as purchasing a toothbrush every 6 months or purchasing a software license every year.

  An example of the result of specifying the appearance pattern is shown in FIG. For example, if the service usage log DB 15 includes a chocolate item ID, based on the corresponding time stamp information, the pattern in which the chocolate appears in the service usage log DB 15 is a specific time type or a constant span type And the specified pattern type and its appearance pattern are stored.

  The specification of the pattern type can be realized by a known technique. For the analysis of specific period type consumption patterns, there is a method of directly analyzing the correlation with the appearance time of the target item, a method of frequency analysis, etc., and a method of analyzing the correlation with the period for items with periodicity. It can be realized using. In addition, the constant span type consumption pattern can be realized by frequency analysis of the item consumption interval for each user or a method of directly analyzing the purchase interval. The specific calculation method of frequency analysis is described in, for example, “M. Vlachos, 2 others,“ Identifying Similarities, Periodicities and Bursts for Online Search Queries ”, Proc. SIGMOD '04”.

  The first edge weight calculation unit 19 calculates the weight for each edge in consideration of the freshness of the history based on the history information with the time stamp acquired from the graph construction unit 16, and assigns the history weight to the history information. This is a functional unit that is returned to the graph construction unit 16 in the assigned form.

For example, the following equation (1) is used to calculate the edge weight W (t) at the time t so that the weight for the edge is attenuated according to the elapsed time from the time stamp time.

Note that T ₀ is an initial value of the weight, t ₀ is a history occurrence time (time stamp time), and t _1/2 is a set value indicating the degree of attenuation.

  The second edge weight calculation unit 20 acquires history information from the graph construction unit 16, acquires appearance pattern identification result information from the appearance pattern DB 18, and records each record of history information based on these two pieces of information. Is a functional unit that performs weighting in consideration of the appearance pattern and returns it to the graph construction unit 16 in the form of the weighting.

  Specifically, for an item corresponding to one of the appearance patterns, the edge weight (weight attenuated according to the elapsed time) related to the item is canceled and a new weight is given.

  For example, assuming each information of FIGS. 2 to 5 and analyzing on February 10, 2014, the condition “Item_keyword = chocolate” on the first line in FIG. 5 and the keyword of the node “Z001” in FIG. So that the weight of the edge associated with such a node is increased. That is, the weight W (t) calculated by the first edge weight calculation unit 19 is changed (corrected) to a weight W (t) ′ larger than that. The corrected weight W (t) ′ is calculated using, for example, an expression obtained by modifying the above expression (1) so that the attenuation portion is not considered.

  The recommendation target specifying unit 21 evaluates the relationship between the target user who recommends the item and the item as the importance on the graph structure acquired from the graph construction unit 16, and the combination of the target user, the item, and its importance This is a functional unit that specifies recommended items based on the above (for example, specifies and identifies the top three importance items as recommended items) and stores them in the recommended item DB 22.

  As a method of calculating the importance, the weight of the target user and each item, the number of steps, a random walk on the graph, and the like are taken into consideration, and the calculation is performed using the above-described RWR algorithm for evaluating the importance between nodes. As for the specific calculation method of the algorithm, for example, “M.Gori, one other,“ ItemRank: A Random-Walk Based Scoring Algorithm for Recommender Engines ”, Proc. IJCAI '07, p.2766-2771” It is described in.

  The recommended item DB 22 is a database that stores the recommended item specified by the recommendation target specifying unit 21 together with the target user and the priority. An example of recommended items is shown in FIG. The combination information of the item ID and priority of the recommended item is stored together with the user ID. The priority is determined according to the importance.

  The recommended item presentation unit 23 accesses the recommended item DB 22 in response to an inquiry from the content providing unit 12, acquires a recommended item for the target user from the recommended item DB 23, and selects one recommended item according to all recommended items or priorities. It is a functional unit that presents the recommended items of the unit to the content providing unit 12.

  The browser function unit 11 is a function unit that provides an interface with a user, receives a user ID, displays information presented by the content providing unit 12 on a screen, and receives a content use request.

  Next, the operation of the information recommendation device 1 will be described. This operation is roughly divided into online processing and offline processing.

  In the online processing, the content providing unit 12 plays a central role, and according to the use of content by the user, the log data resulting from the use is stored in the service use log DB 15 and the recommended items stored in the recommended item DB 22 are presented to the user.

  In the offline processing, first, the appearance pattern DB 18 is constructed using the item usage history and item information stored in the service usage log DB 15 and the item DB 13. Next, the recommended item DB 22 is constructed by further using the appearance pattern specification result stored in the appearance pattern DB 18. This process will be described in detail with reference to FIG.

  First, the appearance pattern specifying unit 17 uses the item usage history stored in the service usage log DB 15 and the item information including the metadata stored in the item DB 13 to generate an item appearance in the entire item usage history. A pattern is specified, and the specified result is stored in the appearance pattern DB 18 (step S1).

  Next, the appearance pattern specifying unit 17 specifies an item appearance pattern for each user using the item usage history and the item information, and further stores the specification result in the appearance pattern DB 18 (step S2).

  Next, based on the user ID and item ID pairs stored in the item usage history, the graph construction unit 16 uses the user and the item as nodes, and connects the nodes on the item usage history with edges. A graph structure is constructed (step S3).

  Next, the graph construction unit 16 inquires of the first edge weight calculation unit 19 about the weight information of each edge considering the freshness of the history with the passage of time, and sets the obtained weight as the edge on the graph structure. (Step S4).

  Next, the graph construction unit 16 inquires of the second edge weight calculation unit 20 about edge weight information that considers the appearance pattern of the item, and determines the weight of the edge related to the corresponding item in consideration of the appearance pattern. The weight is updated (step S5).

  Next, the recommendation target specifying unit 21 calculates the importance of each item based on the graph structure obtained in steps S3 to S5 for the determined user, and determines the recommended item using the importance. Then, the combination information of the item and the priority is stored in the recommended item DB together with the user ID (step S6).

  Finally, it is determined whether or not the processing of step S2 to step S6 has been completed for all users in the item usage history. If not completed, the process returns to step S2, and if completed, this flow ends. (Step S7).

  As described above, according to the present embodiment, if there is an item that matches a specific-period consumption pattern that considers the purchase time of a product or a constant-span consumption pattern that considers a purchase interval, the item includes Since the edge weight on the related graph structure is corrected, an item suitable for the convenience of the user can be recommended. That is, the recommendation of an item that is more meaningful to the user and does not miss the opportunity for item consumption can be realized.

  Finally, the information recommendation device 1 described in the present embodiment can be realized by a computer. Moreover, each operation | movement of the component of this information recommendation apparatus 1 can be constructed | assembled as a program, and it can install and run in the computer utilized as the information recommendation apparatus 1, or can distribute | circulate via a communication network.

DESCRIPTION OF SYMBOLS 1 ... Information recommendation apparatus 11 ... Browser function part 12 ... Content provision part 13 ... Item DB
14 ... Log acquisition unit 15 ... Service usage log DB
16 ... Graph construction unit 17 ... Appearance pattern specifying unit 18 ... Appearance pattern DB
DESCRIPTION OF SYMBOLS 19 ... 1st edge weight calculation part 20 ... 2nd edge weight calculation part 21 ... Recommendation target specific part 22 ... Recommended item DB
23 ... Recommended item presentation unit S1-S7 ... Step

Claims (7)

Storage means for storing a use history in which identification information of an item used by a user and its use time are associated with the identification information of the user;
Obtaining a plurality of usage histories from the storage means, using the usage time included in each usage history, identifying means for specifying a temporal usage pattern for each item for each item or each user;
When there is an item that matches the identification information of the associated item and the identification information of the user with an edge, gives a weight according to the elapsed time from each usage time to each edge, and matches a predetermined usage pattern Generating means for generating a graph structure in which the weight of the edge related to the item is corrected;
Determining means for determining a recommended item based on a weight given to each edge in the graph structure;
An information recommendation device comprising: The generating means includes
2. The information recommendation device according to claim 1, wherein a weight that attenuates according to an elapsed time is given to each edge, and the weight of the edge related to the matching item is corrected to a weight larger than the weight after attenuation. . The determining means includes
The information recommendation device according to claim 1 or 2, wherein the recommended item is determined using a graph structure in which metadata about the item is further connected by an edge. By computer
A storage step of storing in the storage means a use history in which the identification information of the item used by the user and the use time thereof are associated with the identification information of the user;
A specific step of acquiring a plurality of usage histories from the storage means, and specifying a temporal usage pattern in which each item is used for each item or each user, using a usage time included in each usage history;
When there is an item that matches the identification information of the associated item and the identification information of the user with an edge, gives a weight according to the elapsed time from each usage time to each edge, and matches a predetermined usage pattern A generation step of generating a graph structure in which the weight of the edge related to the item is corrected;
Determining a recommended item based on a weight given to each edge in the graph structure;
An information recommendation method comprising: The generating step includes
5. The information recommendation method according to claim 4, wherein a weight that attenuates according to an elapsed time is given to each edge, and an edge weight related to the matching item is corrected to a weight larger than the weight after attenuation. . The determining step includes
6. The information recommendation method according to claim 4, wherein the recommended item is determined using a graph structure in which metadata about the item is further connected by an edge.   An information recommendation program that causes a computer to execute the information recommendation method according to claim 4.

Images