JP5903376B2 - Information recommendation device, information recommendation method, and information recommendation program - Google Patents

Description

  In the present invention, content information and items (collectively referred to as items hereinafter) of a plurality of services and domains (collectively referred to as domains hereinafter) are handled in information recommendation. This is related to information recommendation technology.

  As a conventional technique, there is a method for realizing highly accurate information recommendation by associating metadata assigned to items in a plurality of domains in order to enable item recommendation for a plurality of domains ( For example, see Patent Document 1).

  In addition, for a single domain, clustering that finds a plurality of sub-matrices consisting of user sets and item sets that have similar trends from within the user-item matrix that represents the relationship between users, item purchases, and usage histories There is a research to improve the recommendation accuracy by applying a general collaborative filtering technique using only the partial matrix to which the recommended user belongs. (For example, see Non-Patent Document 1).

JP 2012-150561 A

Bin Xu, Jiajun Bu, Chun Chen, Deng Cai: An Exploration of Improving Collaborative Recommender Systems via User-Item Subgroups, Proceedings of the WWW2012, 2012.

  By using the method of Patent Document 1 described above, information recommendation that handles a plurality of domains can be performed. However, when there is data that does not contribute to the improvement of the recommendation accuracy in the adjacency matrix representing the graph constructed by the recommendation system, there is a problem that the recommendation accuracy decreases.

  In addition, by using the method of Non-Patent Document 1, the recommendation accuracy of a recommendation system that handles a single domain can be expected. However, when clustering is performed for multiple domains, Since a cluster is easily constructed, there are many items that cannot be recommended when recommending across domains, and as a result, there is a problem in that recommendation accuracy decreases.

  In order to solve the above-mentioned problem in Patent Document 1, clustering is performed on the adjacency matrix representing the graph as in Non-Patent Document 1, and a submatrix consisting only of item sets belonging to a cluster strongly associated with the recommended user is used. Can be considered. By doing so, items that do not contribute to recommendation accuracy that easily belong to different clusters can be removed. However, as described above, there is a concern that recommendation accuracy may be reduced due to the loss of information between domains as a result of clustering.

  The present invention has been made in view of the above points, and an object of the present invention is to provide an information recommendation technique that makes it possible to highly accurately recommend items across multiple domains.

In order to solve the above problem, the present invention is an information recommendation device for recommending an item to a recommended user,
By referring to an item history database including a user ID, an item ID, and a domain ID, and using the domain ID information as a constraint, clustering with restrictions on items is performed, and items belonging to a cluster strongly associated with the recommended user Submatrix extraction means for obtaining a set and extracting a submatrix based on the item set;
A recommended item determining means for creating a partial adjacency matrix from the partial matrix, calculating a degree of association between the recommended user and each item using the partial adjacency matrix, and determining a recommended item based on the calculated degree of association; It is comprised as an information recommendation apparatus characterized by providing.

  The sub-matrix extraction unit can be configured to perform the constrained clustering by greatly evaluating the similarity of a set of items belonging to different domains compared to the case of no constraint.

  For example, the sub-matrix extraction unit obtains a constraint vector in which a set having different domain IDs is set to 1 and a set having the same domain ID is set to 0 for the set of items in the item history database, and the constraint vector is used to determine the constraint Perform clustering.

  The recommended item determination means may be configured to calculate the degree of association between the recommended user and each item using an RWR (Random Walk with Restart) algorithm.

  The present invention can also be configured as an information recommendation method executed by an information recommendation device that recommends an item to a recommended user, and an information recommendation program for causing a computer to function as each unit of the information recommendation device.

  According to the present invention, it is decided to perform constrained clustering that leaves information necessary for crossing a plurality of domains, so that it is possible to extract and use only the data that contributes to the recommendation accuracy for items of a plurality of domains. It becomes possible. Therefore, it becomes possible to recommend an item across multiple domains with higher accuracy.

It is a block diagram of the information recommendation apparatus in one embodiment of this invention. It is an image of the table of the item history database in one embodiment of the present invention. It is a flowchart of a process of the information recommendation apparatus in one embodiment of this invention. It is an image of the submatrix in one embodiment of the present invention. It is a flowchart of the constraint vector calculation used by the partial matrix extraction process in one embodiment of this invention. It is a flowchart of the submatrix extraction process in one embodiment of this invention. It is a flowchart of the submatrix normalization process in one embodiment of this invention.

  Embodiments of the present invention will be described below with reference to the drawings. The embodiment described below is only an example, and the embodiment to which the present invention is applied is not limited to the following embodiment.

(Device configuration)
FIG. 1 is a configuration diagram of an information recommendation device according to an embodiment of the present invention.

  The information recommendation apparatus shown in the figure includes a recommended user ID acquisition unit 110, an item history database 120, a partial matrix extraction unit 130, a partial matrix normalization processing unit 140, a recommended item prediction processing unit 150, and an item presentation unit 160. The outline of each functional part is as follows.

  The recommended user ID acquisition unit 110 acquires the recommended user ID and outputs it to the item history database 120 and the recommended item prediction processing unit 150.

  The item history database 120 constitutes storage means for storing an item history table. FIG. 2 shows an image of the table of the item history database in the embodiment of the present invention. As illustrated in FIG. 2, the item history database 120 is a database in which user IDs, item IDs, domain IDs, and evaluation values for items can be referred to.

  The submatrix extraction unit 130 in FIG. 1 calculates a submatrix A, which will be described later, using data in the item history database 120. The partial matrix normalization processing unit 140 normalizes the partial matrix A and creates a partial adjacency matrix. The recommended item prediction processing unit 150 determines a recommended item based on the recommended user ID, the item ID list associated with the recommended user ID, and the partial adjacency matrix. The item presentation unit 160 presents a recommended item that is an output of the recommended item prediction processing unit 150. The partial matrix normalization processing unit 140 and the recommended item prediction processing unit 150 may be collectively referred to as recommended item determination means.

  The information recommendation apparatus according to the present embodiment can be realized, for example, by causing one or a plurality of computers to execute a program describing the processing content described in the present embodiment. That is, the functions of each unit of the information recommendation device (recommended user ID acquisition unit, item history database, partial matrix extraction unit, partial matrix normalization processing unit, recommended item prediction processing unit, item presentation unit) This can be realized by executing a program corresponding to processing executed in each unit using hardware resources such as a CPU, a memory, and a hard disk built in the computer constituting the computer. More specifically, in accordance with a program, data to be calculated is read from the memory, calculation is performed by the CPU, and the process is executed while repeating the operation of storing in the memory. The above-mentioned program can be recorded on a computer-readable recording medium (portable memory or the like), stored, or distributed. It is also possible to provide the program through a network such as the Internet or electronic mail.

  Note that the item history database 120 may not exist in the information recommendation device. For example, the item history database 120 may be provided in an external database server, and the information recommendation device may acquire necessary data by accessing the database server.

(Device operation)
FIG. 3 is a flowchart of the process of the information recommendation device according to the embodiment of the present invention. A process of presenting a recommended item to a recommended user executed by the information recommendation apparatus shown in FIG. 1 will be described along the flowchart of FIG. In the processing of each step shown in FIG. 3, detailed description will be appropriately made with reference to the drawings.

  Step 210) First, as a recommended user ID acquisition process, the recommended user ID acquisition unit 110 acquires the ID of the recommended user and outputs it to the item history database 120 and the recommended item prediction processing unit 150. For example, this process is performed simultaneously with the login process for the service or system used by the user.

  Step 220) As processing of the submatrix extraction unit 130, user ID, item ID, domain ID, and evaluation value data are received from the item history database 120, a submatrix A is generated, and the partial matrix normalization processing unit 140 Output matrix A. FIG. 4 shows an image of the submatrix A in one embodiment of the present invention. As will be described later, the submatrix A is an index of an effective matrix indicating an item ID set included in a cluster having a small distance from a vector related to a recommended user among a plurality of clusters related to items obtained by constrained clustering. Is a matrix.

  Hereinafter, an example of a procedure for constructing the submatrix A in a more detailed embodiment will be described using the flowcharts of FIGS. 5 and 6.

As shown in FIG. 5, first, all user sets U = {u ₁ , u ₂ ,... U _{| U |} } and all item sets I = {i ₁ , i ₂ ,. i _{| I |} } is collected (step 510).

Next, a constraint vector W = (w _{i, j} ) used in constrained clustering is obtained. Specifically, for all the item combinations i, j (i ≠ j) of all the items, 1 is substituted into w _{i, j} for items having different domain IDs on the item history database 120, and combinations that are not 0 is substituted (steps 530 to 537).

Next, constrained clustering is performed using the constraint vector W = (w _{i, j} ) created as described above, and a submatrix A is obtained. This process will be described with reference to FIG.

First, for all items, an item vector x _i having an average of the scores of each user in the item history database 120 as an element is prepared (step 610). K clusters C are prepared for all item vectors (step 620). As the initial value of the cluster, K pieces of x _i are selected at random. The value of parameter K that determines the number of clusters in the clustering result is K = 10 or 50, etc., but it is specified according to the size and use of the data. Is used.

Neighbor calculation to perform clustering by merging x _i to the nearest cluster C _k is performed using the constraint vector w _{i, j} as follows: λ ₁ Σ _xj∈Ck ( W _{i, j} * S im (x _i , x _j )) + λ ₂ Σ _xj∈Ck Sim (x _i , x _j )
Select the one with the maximum value. Where k is a subscript identifying cluster C with 1 ≦ k ≦ K, Sim (x _i , x _j ) represents the similarity between vectors x _i and x _j , and is an arbitrary scale such as Euclidean distance or Jaccad coefficient Can be used. According to the above formula, clustering can be performed by greatly evaluating the degree of similarity of a set of items belonging to different domains as compared to the case of no constraint.

λ _{1 and} λ ₂ are parameters for controlling the degree of influence of similarity with constraints and similarity without constraints, and real numbers such as λ ₁ = 0.85 _and λ ₂ = 0.15 are used. The sum of λ ₁ and λ ₂ is not necessarily 1. For example, if λ ₁ is a large value such as λ ₁ = 0.85 _and λ ₂ = 0.15, the influence of the similarity with restrictions becomes large, and a cluster including items of different domains is easily constructed. By using the partial matrix generated based on the cluster calculated in this way, it is possible to perform highly accurate recommendation across multiple domains.

  The above-mentioned constrained clustering process is performed for all item vectors, and K clusters C are obtained as a result of constrained clustering (steps 620 to 622).

Further, using the item ID set I _a associated with the recommended user a in the item history database 120, the vector of the recommended user is q _a , and its element q _{a, j} is expressed by the following equation:
q _{a, j} = (1 / | I _a |) * Σ _i∈Ia (x _{i, j} / (Σ _{_j} x _{i, j} ))
(Step 630). In the above formula, x _{i, j} is an element of the item vector x _i , and j is a subscript of x _i . x _{i, j} / ( _{Σ_j} x _{i, j} ) means normalization so that the sum of the elements of x _i becomes 1.

Finally, out of K clusters, a matrix having an item ID set included in the cluster C _k having the smallest distance between q _a and the smallest vector as an effective matrix index is defined as a submatrix A = (a _{i, j} ). Output (step 640). That is, an item ID set belonging to a cluster closely related to the recommended user is acquired, and a partial matrix is extracted based on the item ID set. As a standard for the distance between vectors, any standard such as a Euclidean distance or a Jaccard coefficient can be used. The element a _{i, j} of the matrix is Sim (x _i , x _j ).

  Step 230) The submatrix normalization processing unit 140 normalizes the submatrix. This normalization process will be described with reference to the flowchart of FIG.

In the normalization process, first, the partial matrix A = (a _{j, k} ) output from the partial matrix extraction unit 130 is received (step 710).

Next, the value _obtained by dividing each element a _{j, k} of the submatrix A by the sum of the column values is
a _{j, k} = a _{j, k} / Σ _{g = 1} a _{g, k}
Are updated for each element and output as a partial adjacency matrix A (steps 720 to 740). That is, the partial adjacency matrix A is obtained by normalizing the partial matrix A so that the sum of the column values becomes 1.

  Step 240) The recommended item prediction processing unit 150 receives the recommended user ID from the recommended user ID acquisition unit 110, the item ID list associated with the recommended user ID from the item history database 120, and the submatrix normalization processing unit 140. Each partial adjacency matrix is received and an item to be recommended is determined.

For the calculation, Random Walk with Restart (RWR) is used in which the degree of association between the acquired recommended user ID and each item ID is obtained and an item with a high value is recommended. Specifically, p ^(t) is a column vector of the relevance level of an item in the calculation step t, α is a constant satisfying 0 <α <1, A is the partial adjacency matrix, and q is a recommended user item Assuming that a column vector whose value representing n items associated on the history database 120 is 1 / n is represented by the following equation.

p ^{(t + 1)} = (1−α) Ap ^(t) + αq
The calculation is repeated t times on the partial adjacency matrix A, p ^(t) is updated, and a node with a high degree of relevance of the node representing the item is output as a recommended item from the column vector of p ^(t) . Here, t is, for example, 30 times. The number of implementations is not limited to 30 and may be a sufficiently large value.

  Step 250) As the recommended item presentation process, the item presentation unit 160 presents the recommended item that is the output of the recommended item prediction processing unit 150, and the recommendation is completed.

  The present invention is not limited to the above-described embodiments, and various modifications and applications are possible within the scope of the claims.

110 recommended user ID acquisition unit 120 item history database 130 partial matrix extraction unit 140 partial matrix normalization processing unit 150 recommended item prediction processing unit 160 item presentation unit

Claims (6)

An information recommendation device for recommending an item to a recommended user,
By referring to an item history database including a user ID, an item ID, and a domain ID, and using the domain ID information as a constraint, clustering with restrictions on items is performed, and items belonging to a cluster strongly associated with the recommended user Submatrix extraction means for obtaining a set and extracting a submatrix based on the item set;
A recommended item determining means for creating a partial adjacency matrix from the partial matrix, calculating a degree of association between the recommended user and each item using the partial adjacency matrix, and determining a recommended item based on the calculated degree of association; An information recommendation device comprising: The said partial matrix extraction means performs the said restricted clustering by evaluating largely the similarity of the group of the items which belong to a mutually different domain compared with the case where there is no restriction | limiting. Information recommendation device. The sub-matrix extraction unit obtains a constraint vector in which a set having different domain IDs is set to 1 and a set having the same domain ID is set to 0 for the set of items in the item history database, and the constraint clustering is performed using the constraint vector The information recommendation device according to claim 2, wherein: The said recommended item determination means calculates the relevance of the said recommended user and each item using the algorithm of RWR (Random Walk with Restart). Any one of Claim 1 thru | or 3 characterized by the above-mentioned. The information recommendation device described. An information recommendation method executed by an information recommendation device for recommending an item to a recommended user,
By referring to an item history database including a user ID, an item ID, and a domain ID, and using the domain ID information as a constraint, clustering with restrictions on items is performed, and items belonging to a cluster strongly associated with the recommended user A submatrix extraction step of obtaining a set and extracting a submatrix based on the item set;
A recommended item determining step of creating a partial adjacency matrix from the partial matrix, calculating a degree of association between the recommended user and each item using the partial adjacency matrix, and determining a recommended item based on the calculated degree of association; An information recommendation method comprising:   The information recommendation program for functioning a computer as each means of the information recommendation apparatus of any one of Claims 1 thru | or 4.

Images