JP2011113104A - Bidirectional cluster division device, method, and program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a bidirectional cluster division device for simultaneously dividing multi-variable data and sequence data corresponding to the multi-variable data into clusters having common features between the respective variables and between the respective sequence data. <P>SOLUTION: An input means 11 inputs multi-variable data and sequence data corresponding to the multi-variable data. A bidirectional clustering means 12 performs bidirectional clustering to the multi-variable data and the sequence data. The bidirectional clustering means 12 divides the multi-variable data and the sequence data into a plurality of clusters by using evaluation functions indicating whether there are a large number of common features or a small number of the common features between respective variables and between the respective sequence data included in the cluster. <P>COPYRIGHT: (C)2011,JPO&INPIT

Description

  The present invention relates to a bidirectional cluster dividing device, method, and program, and more particularly, a bidirectional cluster dividing device, method, and method for dividing a set of multivariate data into clusters having common characteristics among the data. Regarding the program.

  The clustering technique is a technique for dividing a data set into clusters having common characteristics. Multivariate data is data in which a data point consists of a plurality of variables. A technique for clustering multivariate data for each variable is called one-way clustering. On the other hand, a technique for simultaneously clustering a plurality of variables is called bidirectional clustering (Co-clustering). Non-Patent Documents 1 and 2 are documents in which bidirectional clustering is described.

  Bidirectional clustering is developed especially as a natural language processing technique. In the field of natural language processing, bidirectional clustering is used to simultaneously cluster sentences and words. In bidirectional clustering, multivariate data such as sentences and words is divided into clusters based on co-occurrence information between sentences and words, and clusters in which each subset of sentences and words has a co-occurrence relationship.

  In the field of natural language processing, when sentence and word are clustered without using bidirectional clustering, it is necessary to cluster the sentence and word separately. In sentence clustering, the frequency of words included in each sentence is used as a feature, and cluster division is performed so that sentences having the same feature belong to the same cluster. In word clustering, which sentence contains each word is used as a feature, and cluster division is performed so that words having the same feature are included in the same cluster.

  When unidirectional clustering is used for natural language processing, as described above, sentences are clustered using word features, and words are clustered using sentence features. For this reason, the clustering process becomes redundant. Further, by combining the result of clustering with documents and the result of clustering with words, clustering of both documents and words can be realized. However, in the one-way clustering, since sentences and words are clustered separately, it is difficult to appropriately incorporate the correlation between sentences and words and the co-occurrence relationship into the cluster. On the other hand, in bidirectional clustering, the correlation between documents and words and the co-occurrence relationship can be incorporated into the cluster.

  Patent Document 1 is a document describing a purchase information processing apparatus that extracts clusters from purchase history data of merchandise for each customer. The purchase information processing apparatus of Patent Document 1 includes purchase information generation means and purchase information processing means. The purchase information generating means applies each of the customer and the product in the purchase history data as one item of a row and a column. The purchase information generation means assigns predetermined index values (0 or 1) different from each other to the matrix element of the product with the history of purchase by the customer and the matrix element of the product with no purchase history, Generate.

  The purchase information processing means rearranges the rows in the matrix table based on the sum of the index values for each row, and rearranges the columns based on the sum of the index values for each column. The purchase information processing means rearranges the sum of the index values in ascending order or descending order. The purchase information processing means extracts the clusters defined by the distribution of the index values on the matrix table after the rearrangement. In Patent Document 1, by performing such clustering, it is possible to reduce the amount of calculation and processing time required for cluster extraction of customer information.

  Patent document 2 is a document describing a time-series data processing device for clustering time-series data. The time series data includes at least time information such as processing date and time, customer specifying information, and product specifying information. The time-series data processing apparatus clusters products into a plurality of groups whose purchase customers are similar to each other with respect to time-series data. The time-series data processing apparatus is configured such that, for any two products (products A and B) in the cluster, the number of cases where two products are purchased at the same time, and the number of cases where B is purchased after A is purchased. , Count the number of cases where A is purchased after B is purchased. The time-series data processing apparatus determines the order relationship between the two products from the number of cases counted.

JP 2003-248750 A Japanese Patent Laid-Open No. 9-305571

A Generalized Maximum Entropy Approach to Bregman Co-clustering and Matrix Approximation, A. Banerjee and I. Dhillon and J. Ghosh and S. Merugu and D.S. Modha, KDD2004 Fully Automatic Cross-associations, Deepayan Chakrabarti and Spiros Papadimitriou and Dharmendra S. Modha and Christos Faloutsos, KDD2004 Probabilistic Model-Based Clustering of Multivariate and Sequential Data, Padhraic Smyth, In Proceedings of Artificial Intelligence and Statistics, 1999

  As the information society advances, the amount of accumulated data has become enormous. For example, in the retail industry, a large amount of multivariate data called POS (Point of Sales) data is accumulated. The POS data includes information on which customers have purchased what, when, where and what. Not only multivariate data but also sequence data in which order information is given to each data point is accumulated in an enormous amount. The sequence data is data corresponding to data points, and is data in which information related to two or more keys (attributes) of multivariate data is arranged in time series. When there is sequence data corresponding to the data points of the multivariate data, it is considered preferable to perform clustering in consideration of the sequence data.

  However, Non-Patent Documents 1 and 2 only perform bidirectional clustering on multivariate data, and cannot cluster multivariate data and sequence data simultaneously. Similarly, Patent Document 1 simply performs bidirectional clustering on multivariate data, and cannot perform bidirectional clustering in consideration of sequence data. In addition, Patent Literature 2 merely determines which product was purchased first or at the same time from time-series data for two products belonging to the same cluster after clustering. Clustering considering the above is not performed.

  Here, Non-Patent Document 3 describes a technique for dividing multivariate data and sequence data into clusters. However, clustering in Non-Patent Document 3 is unidirectional clustering. Therefore, in Non-Patent Document 3, multivariate data and sequence data cannot be bi-directionally clustered simultaneously.

  The present invention is a bi-directional cluster division apparatus, method, and method capable of simultaneously dividing multivariate data and sequence data corresponding to the multivariate data into each variable and a cluster having a common feature among the sequence data, and The purpose is to provide a program.

  In order to achieve the above object, the present invention performs multi-variate data and sequence data corresponding to the multivariate data, and performs bi-directional clustering on the multivariate data and the sequence data. Bidirectional clustering means for dividing multivariate data and the sequence data into a plurality of clusters using an evaluation function that indicates whether there are many or less common features between each variable included in the cluster and between the sequence data A bi-directional cluster partitioning device is provided.

  The present invention includes a step of inputting multivariate data and sequence data corresponding to the multivariate data, bi-directional clustering is performed on the multivariate data and the sequence data, and the multivariate data is included in the cluster. There is provided a bidirectional cluster dividing method including a step of dividing into a plurality of clusters using an evaluation function indicating whether there are many or less common features between each variable and between sequence data.

  The present invention provides a computer that inputs multivariate data and sequence data corresponding to the multivariate data, and performs bi-directional clustering on the multivariate data and the sequence data, and the multivariate data is clustered. A program for executing a process of dividing into a plurality of clusters using an evaluation function that indicates whether there are many or less common features between each of the variables included in the sequence data and between the sequence data.

  The present invention accepts a request for content from a user and stores request request means for storing the user who transmitted the request and the requested content in the user request storage unit, and the content requested by the user to the content of the advertiser to the user. Content distribution means for adding and transmitting an advertisement including a mechanism for requesting the user, and based on the information stored in the user request storage unit, the user and the advertisement are used as variables, and the user takes the content of the advertiser from the advertisement. Is generated to generate multivariate data indicating whether or not a request has been made, and to generate sequence data corresponding to the multivariate data, in which the requests transmitted until the user requests the content of the advertiser are arranged in time series Generating means, the multivariate data and the sequence An evaluation function that performs bi-directional clustering on the data and indicates whether the multivariate data and the sequence data have many or few common features between the variables included in the cluster and between the sequence data Bi-directional clustering means for dividing the cluster into a plurality of clusters and outputting a bi-directional clustering result; and an advertisement selecting means for determining an advertisement to be added to the content by the content distribution means based on the bi-directional clustering result; An advertisement distribution system is provided.

  The present invention accepts a request for content from a user, stores a request requesting user and the requested content in a user request storage unit, the content requested by the user, the content of the advertiser to the user A content distribution step of adding and transmitting an advertisement including a mechanism for requesting a request, and using the information stored in the user request storage unit as a variable between the user and the advertisement, Is generated to generate multivariate data indicating whether or not a request has been made, and to generate sequence data corresponding to the multivariate data, in which the requests transmitted until the user requests the content of the advertiser are arranged in time series Generating step, said multivariate data and previous Bidirectional clustering is performed on sequence data, and the multivariate data is divided into a plurality of evaluation functions using an evaluation function that represents whether there are many or few features common between each variable included in the cluster and between the sequence data. There is provided an advertisement distribution method that includes a bidirectional clustering step of dividing a cluster and outputting a bidirectional clustering result, and an advertisement selection step of determining an advertisement to be added to the content based on the bidirectional clustering result.

  The present invention accepts a request for content from a user to a computer, stores the request requesting user and the requested content in a user request storage unit, and advertises the user with the content requested by the user. Based on content distribution processing that adds and transmits an advertisement including a mechanism for requesting the main content, and information stored in the user request storage unit, the user and the advertisement are variables, and the user advertises from the advertisement. In addition to generating multivariate data indicating whether or not the main content has been requested, corresponding to the multivariate data, sequence data in which requests transmitted until the user requests the advertiser content is arranged in time series. Data generation processing to be generated and the multivariate data Bidirectional clustering is performed on the sequence data, and the multivariate data is divided into a plurality of evaluation functions that indicate whether there are many or few features common to each variable included in the cluster and between the sequence data. There is provided a program for executing a bi-directional clustering process for dividing an image into two clusters and outputting a bi-directional clustering result and an ad selection process for determining an advertisement to be added to the content based on the bi-directional clustering result.

  The present invention collects sales information including information that a customer has purchased a product, and based on the collected sales information, the customer and the product are used as variables, and the multivariate indicating whether or not the customer has purchased the product. Generating data, and corresponding to the multivariate data, data generating means for generating a sequence data in which histories about the purchase of a product by a customer are arranged in time series, and the multivariate data and the sequence data Bidirectional clustering is performed on the multivariate data and the sequence data by using an evaluation function that indicates whether there are many or few common features between each variable and sequence data included in the cluster. And a bi-directional clustering means for outputting bi-directional clustering results and a customer based on the bi-directional clustering results. Providing products recommendation system and a recommendation item list generating means for determining a product to be.

  The present invention collects sales information including information that a customer has purchased a product, and based on the collected sales information, the customer and the product are used as variables, and the multivariate indicating whether or not the customer has purchased the product. A data generation step for generating data, and generating data corresponding to the multivariate data, in which the history regarding the purchase of the product by the customer is arranged in time series, and the multivariate data and the sequence data Bidirectional clustering is performed on the multivariate data and the sequence data by using an evaluation function that indicates whether there are many or few common features between each variable and sequence data included in the cluster. A bi-directional clustering step that outputs a bi-directional clustering result, and based on the bi-directional clustering result, To provide products recommendation method and a recommendation item list generation step of determining a commodity to be recommended to the customer.

  The present invention collects sales information including information that a customer has purchased a product on a computer, makes a variable between the customer and the product based on the collected sales information, and determines whether or not the customer has purchased the product. Data generation processing for generating multivariate data to be generated and generating sequence data in which histories relating to the purchase of a product by a customer are arranged in time series corresponding to the multivariate data, and the multivariate data and the sequence Bidirectional clustering is performed on the data, and the evaluation function that indicates whether the multivariate data and the sequence data have many or few common features between each variable included in the cluster and between the sequence data A bi-directional clustering process that divides the data into a plurality of clusters and outputs a bi-directional clustering result; and based on the bi-directional clustering result. Stomach, to provide a program to be executed by the recommendation item list generation process to determine the items to be recommended to the customer.

  The present invention collects failure information including the vehicle type and failure location of the vehicle, and based on the collected failure information, the vehicle type and the region are variables, and whether or not a failure has occurred in the region for the vehicle type. Data generating means for generating sequence data in which the history of failure locations that occurred in the vehicle model in the past is arranged in time series corresponding to the multivariate data, and the multivariate data And multi-variate data and the sequence data are evaluated with respect to each of the variables included in the cluster and between the sequence data. A bidirectional clustering means for dividing the cluster into a plurality of clusters using a function and outputting a bidirectional clustering result, and based on the bidirectional clustering result Occurrence of a failure to provide a failure prediction system comprising a failure prediction candidate list generating unit to estimate a region expected for vehicles.

  The present invention collects failure information including the vehicle type and failure location of the vehicle, and based on the collected failure information, the vehicle type and the region are variables, and whether or not a failure has occurred in the region for the vehicle type. A data generation step for generating multivariate data indicating a sequence data in which a history of fault locations that occurred in the vehicle in the past is arranged in time series corresponding to the multivariate data; and the multivariate data And multi-variate data and the sequence data are evaluated with respect to each of the variables included in the cluster and between the sequence data. A bi-directional clustering step of dividing a multi-cluster using a function and outputting a bi-directional clustering result; Zui and provides a failure prediction method and a failure prediction candidate list generating step to estimate the local occurrence of the failure relative to vehicle type is expected.

  The present invention collects failure information including the vehicle type and failure location of the vehicle in a computer, and based on the collected failure information, the vehicle type and the region are variables, and a failure has occurred in the region for the vehicle type. A data generating process for generating multivariate data indicating whether or not, and corresponding to the multivariate data, generating a sequence data in which histories of fault locations that have occurred in the past in the vehicle type are arranged in time series, and Whether bi-directional clustering is performed on multivariate data and sequence data, and the multivariate data and the sequence data have many or less common features between each variable included in the cluster and between the sequence data A bi-directional clustering process for dividing the cluster into a plurality of clusters using an evaluation function representing Based on the results, the occurrence of a fault with respect to vehicles to provide a program for executing the failure prediction candidate list generation process to estimate an area to be predicted.

  The bidirectional cluster dividing apparatus, method, and program according to the present invention simultaneously divide multivariate data and sequence data corresponding to the multivariate data into clusters having characteristics common to each variable and between the sequence data. can do.

The block diagram which shows the bidirectional | two-way cluster division | segmentation apparatus which concerns on 1st Embodiment of this invention. The flowchart which shows the operation | movement procedure of a bidirectional | two-way cluster division | segmentation apparatus. The figure which shows an example of input data. The figure which shows input data in a table format (matrix format). The figure which shows the result of an initial clustering. The figure which shows the clustering result finally obtained. (A) shows multivariate data, (b) is a figure which shows a clustering result. (A) shows multivariate data to which sequence data is added, and (b) shows a clustering result. The block diagram which shows the advertisement delivery system which concerns on 2nd Embodiment of this invention. The block diagram which shows a user terminal. The block diagram which shows a Web server. The flowchart which shows the procedure of a bidirectional | two-way clustering process. The figure which shows the input data of a bidirectional | two-way clustering means. The figure which shows input data in a table format (matrix format). The figure which shows the clustering result of a bidirectional | two-way clustering means. The flowchart which shows the procedure of an advertisement delivery process. The figure which shows an advertisement delivery candidate. The figure which shows the web page to which the web advertisement was added. The figure which shows the goods recommendation system which concerns on 3rd Embodiment of this invention. The flowchart which shows the operation | movement procedure of goods recommendation. The figure which shows the input data of a bidirectional | two-way clustering means. The figure which shows a bidirectional | two-way clustering result. The figure which shows a recommendation goods list. The block diagram which shows the failure prediction system which concerns on 4th Embodiment of this invention. The flowchart which shows the operation | movement procedure of failure prediction. The figure which shows the input data of a bidirectional | two-way clustering means. The figure which shows a bidirectional | two-way clustering result. The figure which shows a failure prediction candidate list | wrist. The block diagram which shows the outline of the bidirectional | two-way cluster division | segmentation apparatus of this invention.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. FIG. 1 shows a bidirectional cluster dividing device according to a first embodiment of the present invention. The bidirectional cluster dividing device 100 includes an input unit 101, a bidirectional clustering unit 102, a cluster number calculation unit 103, and an output unit 104. The function of each means in the bidirectional cluster dividing device 100 can be realized by a computer reading and executing a predetermined program.

  The input means 101 inputs multivariate data and sequence data. Multivariate data is data having two or more attributes as variables. The sequence data is data corresponding to multivariate data, and is data in which information related to two or more keys (attributes) of the multivariate data is arranged in time series. The multivariate data is, for example, data indicating whether or not the customer has purchased the product, with the customer and the product as variables. The sequence data is, for example, history data in which histories related to the fact that the customer has purchased the product so far are arranged in time series corresponding to the data point that the customer has purchased the product.

  The bidirectional clustering means 102 performs bidirectional clustering on the input data. The bidirectional clustering means 102 divides the multivariate data into a plurality of clusters using the evaluation function. The evaluation function is a function that represents whether the multivariate data has many or few common features between the variables included in the cluster and between the sequence data. For example, if the bi-directional clustering means 102 is a function whose value decreases as the common features of the evaluation function increase, the clustering is performed so that the sum of the evaluation function values calculated for each cluster is small. . The output unit 104 outputs a bidirectional clustering result.

  The cluster number calculation means 103 determines the number of cluster divisions in bidirectional clustering. The cluster number calculation means 103 outputs a predetermined initial value as the number of cluster divisions in the initial clustering. In the initial clustering, the bidirectional clustering means 102 divides input data into a predetermined number of initial clusters. When the bidirectional clustering means 102 performs clustering, the cluster number calculating means 103 determines whether or not to increase the number of cluster divisions based on the value of the evaluation function. When the cluster number calculation unit 103 increases the cluster division number, the bidirectional clustering unit 102 performs the cluster division again with the cluster division number.

FIG. 2 shows an operation procedure. The input means 101 inputs multivariate data and sequence data (step A1). FIG. 3 shows an example of input data. In this example, the multivariate data is data representing who bought which product. There are two variables of multivariate data: “customer” and “product”. Data of the day of the week of product purchase (sequence data) is added to each data of the multivariate data. The sequence data is expressed as y _ik . The sequence data y _jk is data in which the days of purchase of the product k in the past by the customer j are arranged in time series.

Information that a customer has purchased a product can be obtained every predetermined period. The predetermined period is, for example, one month. In FIG. 3, there are two data that customer B purchased product 2, because the period during which customer B purchased product 2 is different. For example, one of the two purchase data corresponds to data that the customer B purchased the product 2 last month, and the other corresponds to data that the customer B purchased the product 2 last month. The sequence data y ¹ _2B represents the purchase day history of the month before the customer B purchased the product 2, and y ² _2B represents the purchase day history of the month before the customer B purchased the product 2.

  FIG. 4 shows the input data in a table format (matrix format). When the input data shown in FIG. 3 is represented by a matrix, it becomes as shown in FIG. A matrix of input data is represented by D. The rows of the matrix D represent customers and the columns represent products. Each element of the matrix D takes a value of 0 or 1. 0 indicates that no product has been purchased, and 1 indicates that a product has been purchased. The sequence data is added to a data point indicating that a customer has purchased a product. The sequence data is not limited to one for one data point, and a plurality of sequence data may correspond to one data point.

    The bidirectional clustering means 102 receives the cluster division number for each variable of the multivariate data from the cluster number calculation means 103. For example, when there are two variables, the bidirectional clustering unit 102 receives the cluster division numbers k and l of each variable from the cluster number calculation unit 103. The bidirectional clustering means 102 performs bidirectional clustering on the multivariate data in consideration of the sequence data (step A2). Bidirectional clustering means 102 divides the multivariate data into k × l clusters. For example, the bidirectional clustering means 102 receives k = 2 and l = 2 as initial values of the number of cluster divisions, and divides the multivariate data into four clusters.

  The bi-directional clustering means 102 performs clustering using the evaluation function. As the evaluation function, a function for calculating the degree to which the data belonging to each cluster does not have a common feature is used. When the input data is bi-directionally clustered, the value of the evaluation function becomes smaller as the data belonging to each cluster has a common feature. Conversely, the value of the evaluation function increases as the data belonging to each cluster does not have a common feature. The bidirectional clustering means 102 performs cluster division so as to reduce the evaluation function.

式１にて、Ｃ（Ｄｉｊ）は、評価関数を用いて計算されるクラスタＤｉｊのコストを表す。コストＴは、各クラスタＤｉｊのコストの総和である。 An evaluation function used in bidirectional clustering considering sequence data will be described. The divided cluster is represented by Dij (i = 1 to k, j = 1 to k). The cost of clustering is defined by the following formula 1.

In Expression 1, C (Dij) represents the cost of the cluster Dij calculated using the evaluation function. The cost T is the sum of the costs of each cluster Dij.

ここで、ｕは、多変量データが取る値である。図４の例では、ｕは、０又は１を取る。ｎ_ｕは、クラスタＤｉｊに属するｕの値の個数を表す。ｎ（Ｄｉｊ）は、クラスタに属するデータ点の数である。すなわち、

である。なお、式２において、ｎ_ｕ（Ｄｉｊ）＝０のときは、

と定義する。 For the cost, a standard called MDL (Minimum Description Length) is used. The cost C (Dij) of each cluster is defined by the following formula 2.

Here, u is a value taken by the multivariate data. In the example of FIG. 4, u takes 0 or 1. n _u represents the number of values of u belonging to the cluster Dij. n (Dij) is the number of data points belonging to the cluster. That is,

It is. In Equation 2, when n _u (Dij) = 0,

It is defined as

ここで、｜Ｄｉｊ｜は、クラスタＤｉｊに属するシーケンスデータの総数を表す。ｍは、ｌｏｇの底である。＾θは、ｙ（Ｄｉｊ）をモデルで表すときのパラメータである。モデルには、シーケンスデータをモデル化する方法として広く利用されているＨＭＭ（Hidden Markov Model）やMarkov Model等の確率モデルを用いることができる。Ｒは、＾θに含まれるパラメータの数を表す。 The function defined by Equation 2 corresponds to the evaluation function. In Equation 2, the value of the first term becomes smaller as the similarity of the multivariate data included in the cluster Dij is higher, and the second term (cost DL (y (Dij))) is the sequence data included in the cluster Dij. The value decreases as the degree of similarity increases. The cost DL (y (Dij)) is defined by the following formula 3.

Here, | Dij | represents the total number of sequence data belonging to the cluster Dij. m is the bottom of the log. ^ Θ is a parameter when y (Dij) is represented by a model. As the model, a probability model such as HMM (Hidden Markov Model) or Markov Model widely used as a method for modeling sequence data can be used. R represents the number of parameters included in ^ θ.

  The cost C (Dij) represents whether the multivariate data and sequence data included in the cluster Dij have many common features or few features. The smaller the value of the cost C (Dij), the more common features, and the larger the value, the fewer common features. When all the multivariate data to which the cluster Dij belongs have the same value, the value of the first term in Equation 2 is 0. In that case, the cost C (Dij) is determined only by DL (y (Dij)). For example, in FIG. 4, the cost of a cluster composed of only u = 1 data points is determined only by the value corresponding to the similarity of the sequence data of the data points belonging to the cluster. Note that the cost of a cluster composed only of u = 0 data points is 0 because there is no sequence data.

  When the bi-directional clustering means 102 performs cluster division, the cluster number calculating means 103 determines whether or not to increase the number of clusters using the cluster division result and the evaluation function (step A3). For example, the cluster number calculation unit 103 determines whether or not the value of the cost T defined by Equation 1 exceeds a predetermined threshold value. The cluster number calculation means 103 determines to increase the number of clusters when the value of the cost T exceeds the threshold value.

  The method for determining whether or not to increase the number of clusters in the cluster number calculation means 103 is not particularly limited to the above. For example, the determination may be made as follows. Any one data point is removed from the multivariate data and the sequence data belonging to the cluster Dij. Let D'ij be the cluster from which one data point is removed. The cluster number calculation means 103 calculates the costs before and after removing the data points, C (Dij) and C (D′ ij), and compares the two. The cluster number calculation means 103 determines to increase the number of clusters when there is a data point where C (Dij)> C (D′ ij).

  When the cluster number calculating unit 103 determines to increase the number of clusters, the cluster number calculating unit 103 notifies the bidirectional clustering unit 102 of the increased number of clusters. For example, the cluster number calculation unit 103 notifies the bidirectional clustering unit 102 of k + 1 and l, k and l + 1, or k + 1 and l + 1 as the new number of clusters, where k and l are the current number of clusters. Thereafter, the process returns from step A3 to step A2, and the bidirectional clustering means 102 divides the input data into the number of notified clusters. By repeating step A2 and step A3, it is possible to obtain an appropriate number of clusters.

  When the output unit 104 determines in step A3 that the cluster number calculation unit 103 does not increase the number of clusters, the output unit 104 outputs the result of the bilateral clustering performed by the bidirectional clustering unit 102 (step A4). The output unit 104 displays, for example, information on data points belonging to each cluster on an output device such as a display for each cluster Dij obtained by cluster division.

FIG. 5 shows the result of the initial clustering. The bidirectional clustering means 102 divides the input data (FIG. 4) into clusters having the initial number of clusters (k = 2, l = 2), thereby obtaining four clusters D11, D12, D21, and D22 shown in FIG. It is done. For each cluster, calculating the cost:
C (D11) = 4log (6/4) + 2log (6/2) + DL (y ¹ _1A , y ¹ _2B , y ² _2B ) = 1.66 + DL (y ¹ _1A , y ¹ _2B , y ² _2B )
C (D12) = 52 log (54/52) +2 log (54/2) + DL (y ¹ _5A , y ¹ _28B ) = 3.72 + DL (y ¹ _5A , y ¹ _28B )
C (D21) = 7log (9/7) + 2log (9/2) + DL (y ¹ _1D , y ¹ _2E ) = 2.07 + DL (y ¹ _1D , y ¹ _2E )
C (D22) = 78 log (81/78) +3 log (81/3) + DL (y ¹ _5D , y ² _28E , y ¹ _30C ) = 5.57 + DL (y ¹ _5D , y ² _28E , y ¹ _30C )
It becomes. The total cost T is
T = C (D11) + C (D12) + C (D21) + C (D22)
= 13.02 + DL (y ¹ _1A , y ¹ _2B , y ² _2B ) + DL (y ¹ _5A , y ¹ _28B ) + DL (y ¹ _1D , y ¹ _2E ) + DL (y ¹ _5D , y ² _28E , y ¹ _30C )
It becomes.

FIG. 6 shows the finally obtained clustering result. By repeatedly performing steps A2 and A3, the number of cluster divisions in the “customer” direction is 2, and the number of cluster divisions in the “product” direction is 3. Finally, the six clusters D11 to D13 shown in FIG. , D21 to D23 are obtained. When the cost of each cluster is calculated for D11 to D13 and D21 to D23 shown in FIG.
C (D11) = 1 log (4/1) +3 log (4/3) + DL (y ¹ _1A , y ¹ _5A , y ¹ _1D ) = 0.997 + DL (y ¹ _1A , y ¹ _5A , y ¹ _1D )
C (D12) = 0
C (D13) = 0
C (D21) = 0
C (D22) = 3 log (9/3) +6 log (9/6) + DL (y ¹ _2B , y ² _2B , y ¹ _28B , y ¹ _2E , y ² _28E , y ¹ _2C , y ¹ _30C ) = 2. 48 + DL (y ¹ _2B , y ² _2B , y ¹ _28B , y ¹ _2E , y ² _28E , y ¹ _2C , y ¹ _30C )
C (D23) = 0
It becomes. The total cost T is
T = ΣC (Dij) = 3.46 + DL (y ¹ _1A , y ¹ _5A , y ¹ _1D ) + DL (y ¹ _2B , y ² _2B , y ¹ _28B , y ¹ _2E , y ² _28E , y ¹ _2C , y ¹ _30C )
It becomes.

  When the cost T in the clustering result shown in FIG. 5 is compared with the cost T in the clustering result shown in FIG. 6, it is confirmed that the value of the evaluation function is reduced except for the DL value (similarity of sequence data). it can. That is, by performing bi-directional clustering based on the evaluation function, multivariate data and sequence data corresponding to the multivariate data can be divided into clusters having features common to the respective variables and between the sequence data. Note that the cost T is not limited to the above, and may include other costs necessary for bidirectional clustering.

  As a comparative example, consider bidirectional clustering that does not consider sequence data. Consider bivariate data as multivariate data. One variable is the customer and the other is the product. FIG. 7A shows multivariate data. The customer takes A, B, and C values, and the product takes 1, 2, and 3 values. The value of the multivariate data represents whether or not the customer has purchased the product. For example, when customer A purchases product 1, the value of the data point corresponding to customer A and product 1 is 1.

  When the multivariate data shown in FIG. 7A is clustered in both directions of the customer and the product, the clustering result shown in FIG. 7B is obtained. In this case, the number of cluster divisions is four. By performing bi-directional clustering on multivariate data, a cluster consisting of data points that customers A and C purchase products 1 and 3 and a cluster consisting of data points that customer B purchases product 2 Is obtained. From this clustering result, it can be seen that customers A and C have characteristics common to products 1 and 3, and customer B has characteristics common to product 2.

  FIG. 8 shows an example of bidirectional clustering of multivariate data and sequence data. FIG. 8A shows multivariate data to which sequence data is added. The sequence data includes, for example, data indicating the day of the week on which the customer has purchased the product. The sequence data is attached to a data point indicating that a customer has purchased a product, that is, a data point having a value of 1.

  When the multivariate data shown in FIG. 8A is bi-directionally clustered in consideration of not only customers and products but also sequence data, the clustering result shown in FIG. 8B is obtained. In this case, the number of cluster divisions is 6. By performing bi-directional clustering in consideration of sequence data, a cluster consisting of data points that customers A and C purchase product 1 and a cluster consisting of data points that customers A and C purchase product 3 , A cluster consisting of data points that customer B purchases product 2 is obtained.

  From the bidirectional clustering result shown in FIG. 8B, it can be seen that the customers A and C have purchased the product 1 with the same purchase day history. Further, it can be read that the customers A and C purchase the product 3 with the same purchase day history. Customers A and C both purchase product 1 and product 3, but product 1 and product 3 are not classified into the same cluster, so product 1 and product 3 have different purchase day histories. Can read that. That is, it can be read that the customers A and C do not purchase the product 1 at the same day interval as the product 3. Regarding the product 2, it can be read that the day of the week history of the purchase of the product 2 by the customer B is different from the purchase day history of the products 1 and 3.

  In the present embodiment, the bidirectional clustering means 102 performs bidirectional clustering on multivariate data and sequence data. As an evaluation function, multi-variate data and multivariate data are obtained by performing bi-directional clustering using an evaluation function that indicates whether there are many or few common features between each variable included in the cluster and between sequence data. Corresponding sequence data can be simultaneously divided into clusters having common features between the variables and between the sequence data.

  In the present embodiment, the bidirectional cluster dividing device 100 includes the cluster number calculating unit 103. The cluster number calculation means 103 determines whether or not the clustering result is appropriate based on the evaluation function, and increases the number of cluster divisions in order to obtain a better clustering result. In clustering, it is often not known in advance how many clusters should be divided. In the present embodiment, the cluster number calculation unit 103 dynamically determines the number of clusters, so that even when it is not known in advance how many clusters should be divided, the multivariate data is converted into an appropriate number of divisions. Can be divided into clusters.

  FIG. 9 shows an advertisement distribution system according to the second embodiment of the present invention. The advertisement distribution system includes a bidirectional cluster dividing device 100 and a Web server 300. The configuration of the bidirectional cluster dividing device 100 is the same as the configuration of the bidirectional cluster dividing device in the first embodiment shown in FIG. The Web server 300 is connected to the user terminal 200 via a network such as the Internet 400. The user terminal 200 provides an interface such as input / output to the user. The user terminal 200 is, for example, a personal computer or a portable information terminal device.

  When a user requests web content, the advertisement distribution system adds an advertisement to the content requested by the user and distributes the content to the user. The advertisement includes a mechanism for allowing the user to request the content of the advertiser. More specifically, the advertisement includes a link of a site that the advertiser wants to guide, and the user can click on the advertisement so that the user can visit a site such as the advertiser. The advertiser includes, for example, a link to a web page on which detailed information on products or services is posted, and guides the user to the web page.

  Here, even if an advertisement is delivered to the user along with the content, if the advertisement is different from the user's preference, the user is unlikely to click on the advertisement and can be directed to a site that the user wants to visit. Less likely. It is difficult for the advertiser to obtain the effect of advertisement distribution unless the user clicks on the advertisement. Therefore, in the advertisement distribution system, it is important to accurately predict an advertisement that matches the user's preference.

  The Web server 300 provides multivariate data and sequence data to the bidirectional cluster dividing device 100. The bidirectional cluster dividing device 100 performs bidirectional clustering on multivariate data and sequence data. The Web server 300 receives the bidirectional clustering result from the bidirectional cluster dividing device 100. The Web server 300 uses the bidirectional clustering result to distribute an advertisement corresponding to the user's preference to the user who requested the Web content. This embodiment applies to the field of collaborative filtering and collaborative filtering.

  FIG. 10 shows the user terminal 200. The user terminal 200 includes a content request unit 201 and a content display unit 202. The content request unit 201 requests the Web server 300 for content that the user desires to browse. The content display unit 202 acquires the content requested by the user from the Web server 300 and displays it. The function of each unit in the user terminal 200 can be realized by the computer operating according to a predetermined program.

  For example, when the user desires sports content, the content request unit 201 requests the Web server 300 for sports content. Further, when the user clicks on an advertisement distributed along with the content, the content request unit 201 requests the Web server 300 for content corresponding to the advertisement.

  FIG. 11 shows the Web server 300. The Web server 300 includes a content distribution unit 301, a user request storage unit 302, a content storage unit 303, an advertisement selection unit 304, an advertisement storage unit 305, a request reception unit 306, a clustering control unit 307, an output device 308, an input device 309, and And a clustering result storage unit 310. The function of each unit in the Web server 300 can be realized by a computer operating according to a predetermined program.

  The request receiving unit 306 receives a request from the user. The request from the user includes a request for requesting acquisition of a desired Web page and a request for requesting acquisition of a Web page corresponding to the Web advertisement. The user request storage unit 302 stores information related to requests from users. The request receiving unit 306 stores, for example, the user name, the request content, and the request time in the user request storage unit 302.

  The content storage unit 303 stores content to be distributed to the user. The advertisement storage unit 305 stores web advertisements. The content distribution unit 301 acquires the content requested by the user from the content storage unit 303 and distributes the content to the user. At that time, the content distribution unit 301 adds the Web advertisement stored in the advertisement storage unit 305 to the content and distributes the content to the user. The content distribution unit 301 may transfer the request to an external server when the content requested by the user is not in the content storage unit 303. Further, when the content requested by the user is a Web page corresponding to the advertisement, the content distribution unit 301 may not add the Web advertisement to the content.

  The clustering control unit 307 also serves as a data generation unit. The clustering control unit 307 performs generation of data to be given to the bidirectional cluster dividing device 100 and control of bidirectional clustering performed by the bidirectional cluster dividing device 100. For example, when the total number of accesses to the Web server 300 exceeds a predetermined threshold, the clustering control unit 307 reads past content visit histories of all users from the user request storage unit 302. Based on the read information, the clustering control unit 307 uses the user and the advertisement as variables, and generates multivariate data indicating whether the user has requested the advertiser's content from the advertisement. Since the user requests the advertiser's content by clicking on the advertisement, the multivariate data represents data indicating which advertisement the user has clicked.

  Further, the clustering control unit 307 generates sequence data corresponding to the multivariate data, in which the requests transmitted until the user requests the content of the advertiser are arranged in time series. Below, the data which arranged the request which the user transmitted in time series are also called content visit history. The clustering control unit 307 passes the multivariate data indicating which user has clicked which advertisement and the content visit history (sequence data) until the advertisement is clicked to the output device 308, and the bidirectional cluster dividing device. Request 100 for bi-directional clustering.

  The interactive cluster dividing device 100 inputs multivariate data indicating which user has clicked which advertisement and the content visit history until the user clicked the web advertisement via the output device 308. The bidirectional cluster dividing device 100 performs bidirectional clustering on the multivariate data and the content visit history until the user clicks on the Web advertisement. The bidirectional cluster dividing device 100 outputs the bidirectional clustering result to the Web server 300.

  The input device 309 inputs the bidirectional clustering result from the bidirectional cluster dividing device 100 and passes it to the clustering result storage unit 310. The clustering result storage unit 310 stores the bidirectional clustering result received from the input device 309. The advertisement selection unit 304 refers to the clustering result storage unit 310 and determines a Web advertisement to be distributed to the user based on the bidirectional clustering result. The advertisement selection unit 304 reads the web advertisement from the advertisement storage unit 305 and gives it to the content distribution unit 301.

  The operation procedure will be described below. The operation of the advertisement distribution system is roughly divided into two types, that is, a bidirectional clustering process and an advertisement distribution process using the bidirectional clustering result. FIG. 12 shows the procedure of bidirectional clustering processing. When the user requests content, the content request unit 201 of the user terminal 200 requests the content from the Web server 300 (step B1). The user is a user whose attribute information is known in advance, and the Web server 300 can determine which user the content request is from.

  The request receiving unit 306 of the Web server 300 receives a request from a user. The request accepting unit 306 stores the user name, request content, and time in the user request storage unit (step B2). Further, the request reception unit 306 passes a request from the user to the content distribution unit 301.

  The content distribution unit 301 reads the content corresponding to the request from the content storage unit 303. Further, the content distribution unit 301 receives a Web advertisement from the advertisement selection unit 304. The content distribution unit 301 adds a Web advertisement to the content read from the content storage unit 303 and transmits it to the user terminal 200 (step B3). The content display unit 202 of the user terminal 200 displays the received content (step B4).

  The clustering control unit 307 determines whether or not the total number of accesses to the Web server 300 has exceeded a predetermined threshold value. When the clustering control unit 307 determines that the total number of accesses exceeds the threshold, the clustering control unit 307 reads the past content visit history of all users from the user request storage unit 302 (step B5). The clustering control unit 307 generates multivariate data indicating which user has clicked which advertisement and content visit history until the user clicked the web advertisement based on the read content visit history. The clustering control unit 307 outputs the generated multivariate data and the content visit history until the user clicks the Web advertisement to the bidirectional cluster dividing device 100 (step B6).

  The bidirectional clustering unit 102 (FIG. 1) of the bidirectional cluster dividing device 100 inputs multivariate data and a content visit history until the user clicks on the Web advertisement via the input unit 101. The bidirectional clustering means 102 performs bidirectional clustering on the multivariate data and the content visit history until the user clicks on the web advertisement (step B7). The procedure of bidirectional clustering is the same as the procedure shown in FIG.

  The bidirectional clustering unit 102 transmits the bidirectional clustering result to the Web server 300 via the output unit 104 (step B8). Upon receiving the bidirectional clustering result, the input device 309 of the Web server 300 stores the received bidirectional clustering result in the clustering result storage unit 310. The clustering result storage unit 310 stores the bidirectional clustering result (step B9).

  FIG. 13 shows input data of the bidirectional clustering means 102. There are two variables of multivariate data: “user” and “Web advertisement”. The content visit history is sequence data in which requests until a user clicks on an advertisement are arranged in time series. The content visit history is, for example, a time series of requests from the first request sent by the user on the day when the user clicks on the advertisement to the request immediately before clicking on the advertisement. Alternatively, the content visit history may be a time series of requests up to 10 before the user clicks on the advertisement. The definition of the content visit history is not particularly limited to the above.

The content visit history (sequence data) is represented by y ⁱ _jk . y ⁱ _jk is sequence data corresponding to the data point that the user k clicked the advertisement j, and is a content visit history in which requests transmitted until the user k clicked the advertisement j are arranged in time series. i represents the number of times the user clicked on the advertisement. For example, y ¹ _jk represents the content visit history when the user k clicks the advertisement j for the first time, and y ² _jk represents the content visit when the user k clicks the advertisement j for the second time. Represents history.

  FIG. 14 shows the input data in a table format (matrix format). When the input data shown in FIG. 13 is represented by a matrix, it becomes as shown in FIG. A matrix of input data is represented by D. The rows of the matrix D represent users, and the columns represent web advertisements. Each element of the matrix D takes a value of 0 or 1. 0 indicates that the user has not clicked on the advertisement, and 1 indicates that the user has clicked on the advertisement. The sequence data is not limited to one for one data point, and a plurality of sequence data may correspond to one data point.

  FIG. 15 shows the clustering result of the bidirectional clustering means 102. The bidirectional clustering means 102 divides input data into six clusters D11 to D13 and D21 to D23 shown in FIG. 15 by performing bidirectional clustering on the multivariate data and sequence data shown in FIG. .

When the cost of each cluster is calculated for D11 to D13 and D21 to D23 shown in FIG.
C (D11) = 1 log (4/1) +3 log (4/3) + DL (y ¹ _1A , y ¹ _5A , y ¹ _1D ) = 0.997 + DL (y ¹ _1A , y ¹ _5A , y ¹ _1D )
C (D12) = 0.0
C (D13) = 0.0
C (D21) = 0.0
C (D22) = 3 log (9/3) +6 log (9/6) + DL (y ¹ _2B , y ² _2B , y ¹ _28B , y ¹ _2E , y ² _28E , y ¹ _2C , y ¹ _30C ) = 2. 48 + DL (y ¹ _2B , y ² _2B , y ¹ _28B , y ¹ _2E , y ² _28E , y ¹ _2C , y ¹ _30C )
C (D23) = 0.0
It becomes. The total cost T is
T = ΣC (Dij) = 3.46 + DL (y ¹ _1A , y ¹ _5A , y ¹ _1D , y ¹ _5D ) + DL (y ¹ _2B , y ² _2B , y ¹ _28B , y ¹ _2E , y ² _28E , y ¹ _2C , y ¹ _30C )
It becomes.

  Subsequently, an advertisement distribution process using the bidirectional clustering result will be described. FIG. 16 shows the procedure of the advertisement distribution process. The content request unit 201 of the user terminal 200 requests content from the Web server 300 (step C1). The request receiving unit 306 receives a request from the user terminal 200. The user request storage unit 302 stores the request received by the request receiving unit 306 (step C2).

  The content distribution unit 301 receives a request from the request reception unit 306. The content distribution unit 301 passes information for identifying the user who transmitted the request, for example, the user name to the advertisement selection unit 304 (step C3). The advertisement selection unit 304 reads the cluster information to which the user belongs from the clustering result storage unit 310 (step C4). In step C4, the clustering result storage unit 310 reads the user name of the user belonging to the cluster to which the user belongs and the information for identifying the Web advertisement clicked by the user of the belonging class.

  The advertisement selection unit 304 determines a web advertisement to be distributed to the user who has requested the content based on the information read in step C4 (step C5). The advertisement selection unit 304 sets an advertisement clicked by a user belonging to the same cluster as an advertisement candidate to be distributed to the user, and determines an advertisement to be distributed to the user from the candidates. In the advertisement determination, the advertisement selection unit 304 determines that the advertisement is preferentially delivered to the user when there is an advertisement that is clicked by another user but is not clicked by the user who has requested the content. To do.

  FIG. 17 shows advertisement distribution candidates. When the cluster division result shown in FIG. 15 is obtained as the cluster division result, advertisement candidates to be distributed to each user are as shown in FIG. In FIG. 17, the order in which the Web advertisement distribution candidates are arranged is the order of priority. For example, consider cluster D11. Referring to FIG. 15, there are two users belonging to this cluster, user A and user D. User A clicks advertisement 1 and advertisement 5, and user D clicks advertisement 1.

  Users belonging to the same cluster have similar preferences regarding Web advertisements and are considered to be interested in Web advertisement groups belonging to the cluster. In addition, since the interactive cluster dividing apparatus 100 performs bidirectional clustering using the sequence data indicating which Web page was accessed and then clicked on the advertisement, users belonging to the same cluster can There are many common features. For this reason, if an advertisement that has been clicked by at least one of the users belonging to the same cluster is distributed to the user who requested the content, the user who received the advertisement is predicted to click the advertisement. it can.

  Since the user belonging to the cluster D11 has clicked on the advertisement 1 and the advertisement 5, the advertisement candidates to be distributed to the user A and the user D are the advertisement 1 and the advertisement 5. Since the user D has never clicked the advertisement 5, the advertisement selection unit 304 makes the priority of the advertisement 5 higher than the priority of the advertisement 1. The advertisement selection unit 304 determines advertisements to be distributed to the user D in the order of the advertisement 5 and the advertisement 1 in accordance with the priority order. Since user A clicks advertisement 1 and advertisement 5, there is no particular priority. The advertisement selection unit 304 may determine, for the user A, any one of the advertisement 1 and the advertisement 5 as an advertisement to be distributed to the user A at random.

  Returning to FIG. 16, when the advertisement selection unit 304 determines an advertisement to be distributed, the advertisement selection unit 304 reads a Web advertisement from the advertisement storage unit 305 and supplies the Web advertisement to the content distribution unit 301. The advertisement selection unit 304 may pass information for identifying the determined Web advertisement to the content distribution unit 301, and the content distribution unit 301 may read the Web content from the advertisement storage unit 305.

  The content distribution unit 301 reads the content requested by the user from the content storage unit 303 (step C6). The content distribution unit 301 adds the Web advertisement determined by the advertisement selection unit 304 to the read content (step C7). The content distribution unit 301 transmits the Web content added with the Web advertisement to the user terminal 200 (Step C8). The content display unit 202 of the user terminal 200 displays the web content including the web advertisement transmitted by the content distribution unit 301 (step C9).

  FIG. 18 shows a Web page to which a Web advertisement is added. The content distribution unit 301 provides an advertisement display area 902 in the web page 901 and embeds a web advertisement in the advertisement display area 902. Web advertisement distribution is not particularly limited to the one described here. For example, the web advertisement may be distributed separately from the web content without embedding the web advertisement in the web content.

  In this embodiment, the bi-directional cluster partitioning device uses multivariate data as data indicating which user clicked which web advertisement, and multi-variate data using the content visit history until the user clicked the web advertisement as sequence data. Bidirectional clustering is performed on the sequence data. The feature of the user also appears in the information on how to click on the Web advertisement in addition to the information on which Web advertisement is requested. In this embodiment, since multivariate data and sequence data are handled simultaneously and bidirectional clustering is performed on them, it can be expected that user characteristics and preferences can be extracted more accurately. Moreover, it can be expected that the user clicks on the advertisement by determining the advertisement to be distributed to the user using the result of such bidirectional clustering.

  FIG. 19 shows a product recommendation system according to the third embodiment of the present invention. The product recommendation system has a server system 600. The server system 600 includes a bidirectional cluster dividing device 100, a data generation unit 601, a recommended product list generation unit 602, and a clustering result storage unit 603. A server system 600 is connected to client systems 501 to 503 via a network 401. The client systems 501 to 503 are sales management systems installed in retail stores, for example. The server system 600 is a central management system that bundles retail store information, and is installed in a data center or the like.

  The client systems 501 to 503 manage sales information of each store. The sales information includes, for example, a customer name, a product name purchased by the customer, and information related to the purchase date and time. The server system 600 collects customer purchase information including data indicating which customers have purchased which products from the client system. The server system 600 performs bidirectional clustering using the collected information. When such information is used as multivariate data, the bidirectional clustering result can be used for retail marketing or the like. The server system 600 determines a product to be recommended to the customer in the future using the bidirectional clustering result. The server system 600 transmits recommended product information to the client systems 501 to 503 of each store.

  The data generation unit 601 collects customer purchase information from the client systems 501 to 503. Based on the collected customer purchase information, the data generation unit 601 uses the customer and the product as variables, and generates multivariate data indicating whether the customer has purchased the product. Further, the data generation unit 601 generates sequence data in which histories related to the purchase of a product by a customer are arranged in time series corresponding to the multivariate data. Here, it is assumed that the sequence data is a purchase day history in which the customer's product purchase days are arranged in time series.

  The configuration of the bidirectional cluster dividing device 100 is the same as the configuration of the bidirectional cluster dividing device in the first embodiment shown in FIG. The bidirectional cluster dividing device 100 performs bidirectional clustering on the multivariate data and sequence data generated by the data generation unit 601. The bidirectional bidirectional cluster dividing device 100 stores the bidirectional clustering result in the clustering result storage unit 603. The recommended product list generation unit 602 generates a list of products recommended to the customer using the clustering result stored in the clustering result storage unit 603.

  FIG. 20 shows an operation procedure. Each of the client systems 501 to 503 transmits customer purchase information to the server system 600 via the network 401 (step D1). The server system 600 receives customer purchase information from each client. The timing at which each client transmits customer purchase information to the server system 600 may be different for each client.

  The data generation means 601 generates multivariate data indicating which customers have purchased which products and purchase day history indicating the history of the days on which the customers have purchased the products. The data generation unit 601 outputs the generated multivariate data and purchase day history to the bidirectional cluster dividing device 100 (step D2).

  Bidirectional clustering means 102 (FIG. 1) of bidirectional cluster dividing apparatus 100 inputs multivariate data and purchase day history via input means 101. The bidirectional clustering means 102 performs bidirectional clustering on the multivariate data and the purchase day history (step D3). The procedure of bidirectional clustering is the same as the procedure shown in FIG. The bidirectional cluster dividing device 100 sends the bidirectional clustering result to the clustering result storage unit 603, and stores the bidirectional clustering result in the clustering result storage unit 603 (step D4).

  The recommended product list generation unit 602 reads the bidirectional clustering result from the bidirectional clustering result storage unit 603, and generates a recommended product list for each customer (step D5). In step D5, the recommended product list generation unit 602 checks, for each cluster, products purchased by at least one of the customers belonging to the cluster. For each customer, the recommended product list generation unit 602 includes, in the recommended product list, products that the customer has not purchased among products purchased by at least one of the customers who belong to the cluster to which the customer belongs.

  The recommended product list generation unit 602 transmits the recommended product list to the client systems 501 to 503 (step D6). The client systems 501 to 503 receive the recommended product list for each customer from the server system 600 (step D7).

  FIG. 21 shows input data of the bidirectional clustering means 102. There are two variables of multivariate data: “customer” and “product”. The purchase day history is sequence data in which, for example, in January, the customer's purchase day history is arranged in time series. FIG. 22 shows the bidirectional clustering result. Assume that the bi-directional clustering means 102 performs bi-directional clustering on the input data shown in FIG. 21 to obtain 2 × 3 = 6 clusters shown in FIG.

  FIG. 23 shows a recommended product list. When the cluster division result shown in FIG. 22 is obtained as the cluster division result, a list of products to be recommended to each customer (recommended product candidates) is as shown in FIG. For example, consider cluster D11. Referring to FIG. 22, there are two customers belonging to this cluster, customer A and customer D. Customer A purchases product 1 and product 5, and customer D purchases product 1.

  In the present embodiment, bi-directional clustering is performed on customers, products, and purchase day histories. By bi-directional clustering, customers who are interested in the same product and have similar purchase day histories of products can be obtained. Can be collected in each cluster. Customers belonging to the same cluster have similar preferences regarding purchased products, and it is considered that there are many common features in the day of the week history of purchasing products. Therefore, if a recommended product related to a product belonging to a cluster is recommended to a customer belonging to the cluster, it can be expected that the customer purchases the product.

  Since the customer belonging to the cluster D11 has purchased the product 1 and the product 5, the recommended product list generation unit 602 selects the product recommended for the customer A and the customer D from the product 1 and the product 5. Choose. Since the customer D has not purchased the product 5, the recommended product list generation unit 602 determines the product recommended for the customer D as the product 5. Since the customer A has already purchased the product 1 and the product 5, the recommended product list generation unit 602 determines that there is no product recommended for the customer A.

  In the present embodiment, the bidirectional cluster dividing device 100 uses multivariate data as data indicating which customers have purchased which products, and multivariate data and sequence data using the history of days of the week when customers have purchased products as sequence data. And bi-directional clustering. In addition to the information on which products have been purchased, the customer characteristics also appear in the information on what day of the week the products have been purchased. In this embodiment, since multivariate data and sequence data are handled simultaneously and bidirectional clustering is performed on them, it can be expected that user characteristics and preferences can be extracted more accurately. Further, by determining a product to be recommended to a customer using the result of performing such bidirectional clustering, a product that the user can expect to purchase can be set as a recommended product.

  FIG. 24 shows a failure prediction system according to the fourth embodiment of the present invention. The failure prediction system has a server system 800. The server system 800 includes a bidirectional cluster dividing device 100, a data generation unit 801, a failure prediction candidate list generation unit 802, and a clustering result storage unit 803. The server system 800 is connected to client systems 701 to 703 via a network 402. The client systems 701 to 703 are installed in, for example, car dealers and repair shops. The server system 800 is a central management system and is installed in a data center or the like.

  The client systems 701 to 703 manage vehicle failure information. The failure information includes a vehicle type and a failure location (failed part). For example, it is assumed that failures have occurred in a plurality of regions for each vehicle type, and the client systems 701 to 703 store failure histories of parts that have failed for each vehicle type. The server system 800 collects failure information from the client systems 701 to 703. The server system 800 performs bidirectional clustering using information collected from the client system. The server system 800 performs failure prediction using the bidirectional clustering result, and transmits the prediction result to the client systems 701 to 703.

  The data generation unit 801 collects failure information for each vehicle type from the client systems 701 to 703. The data generation unit 801 uses the vehicle type and the region as variables based on the collected failure information, and generates multivariate data indicating whether or not a failure has occurred in the vehicle type in the region. Further, the data generation unit 801 generates sequence data in which histories related to the occurrence of a failure in the vehicle type are arranged in time series corresponding to the multivariate data. Here, it is assumed that the sequence data is a failure part history in which parts that have failed in the past are arranged in time series.

  The configuration of the bidirectional cluster dividing device 100 is the same as the configuration of the bidirectional cluster dividing device in the first embodiment shown in FIG. The bidirectional cluster dividing device 100 performs bidirectional clustering on the failure occurrence area for each vehicle type and the failure part history. The clustering result storage unit 803 stores the clustering result of the bidirectional cluster dividing device 100. The failure prediction candidate list generation unit 802 generates a list of parts that are predicted to have a failure in the future for the vehicle type and the region, using the clustering result stored in the clustering result storage unit 803.

  FIG. 25 shows an operation procedure. Each of the client systems 701 to 703 transmits failure information to the server system 800 via the network 402 (step E1). The server system 800 receives failure information from each client. It is assumed that failure information managed by the client systems 701 to 703 has different regions, and the server system 800 can determine the region where the failure has occurred, depending on from which client the failure information is received. Alternatively, the failure information may include information regarding the area. The timing at which each client transmits failure information to the server system 800 may be different for each client.

  The data generation unit 801 generates multivariate data that indicates in which vehicle a failure has occurred in which vehicle type, and a failure component history that indicates a history of components that have previously failed in that vehicle type. The data generation unit 801 outputs the generated multivariate data and the failure part history to the bidirectional cluster dividing device 100 (step E2).

  Bidirectional clustering means 102 (FIG. 1) of bidirectional cluster dividing apparatus 100 inputs multivariate data and a failure part history via input means 101. The bidirectional clustering means 102 performs bidirectional clustering on the multivariate data and the failure part history (step E3). The procedure of bidirectional clustering is the same as the procedure shown in FIG. The bidirectional cluster dividing device 100 sends the bidirectional clustering result to the clustering result storage unit 803, and stores the bidirectional clustering result in the clustering result storage unit 803 (step E4).

  The failure prediction candidate list generation unit 802 reads the bidirectional clustering result from the bidirectional clustering result storage unit 803, and generates a failure prediction candidate list for each vehicle type (step E5). In step E5, the failure prediction candidate list generation unit 802 checks, for each cluster, an area where a failure has occurred in at least one of the vehicle types belonging to the cluster. The failure prediction candidate list generation unit 802 selects, for each vehicle type, a region where a failure has not yet occurred in the customer from among regions where a failure has occurred in at least one of the vehicle types belonging to the cluster to which the vehicle type belongs. Include in the failure prediction candidate list.

  The failure prediction candidate list generation unit 802 transmits the failure prediction candidate list to the client systems 701 to 703 (step E6). The client systems 701 to 703 receive a failure prediction candidate list for each customer from the server system 800 (step E7).

  FIG. 26 shows input data of the bidirectional clustering means 102. There are two variables of the multivariate data: “car type” and “region”. The failure part history is sequence data in which, for example, in a year unit, the history of parts that have failed in the vehicle type is arranged in time series. Alternatively, the failed part history may be a list of past failed parts in which a failure occurred before the failure occurred. FIG. 27 shows the bidirectional clustering result. Assume that the bi-directional clustering means 102 performs bi-directional clustering on the input data shown in FIG. 26 to obtain 2 × 3 = 6 clusters shown in FIG.

  FIG. 28 shows a failure prediction candidate list. When the cluster division result shown in FIG. 27 is obtained as the cluster division result, a list of areas (failure occurrence area candidates) where a failure is predicted to occur in each vehicle type is as shown in FIG. For example, consider cluster D11. Referring to FIG. 27, there are two vehicle types belonging to this cluster, vehicle type A and vehicle type D. Further, the failure of the vehicle type A occurs in the region 1 and the region 5, and the failure of the vehicle type D occurs in the region 1.

  In this embodiment, bi-directional clustering is performed on the vehicle type, region, and failure part history. By performing bi-directional clustering, a failure occurs in the same region, and a vehicle type that has a similar failure equipment history is also obtained. Can be collected in each cluster. The types of vehicles belonging to the same cluster tend to have the same failure occurrence area, and it is considered that the history of parts in which a failure has occurred includes many common features. Therefore, it can be predicted that a vehicle type belonging to a certain cluster will fail in an area belonging to the cluster in the future.

  The failure prediction candidate list generation means 802 selects the failure occurrence region from the regions 1 and 5 because the vehicle type belonging to the cluster D11 has a failure in the region 1 and the region 5. Since the failure has already occurred in the region 1 for the vehicle type D, the failure prediction candidate list generation unit 802 determines the region in which the occurrence of the failure is predicted as the region 5 in the vehicle type D. Since the vehicle type A has already failed in the region 1 and the region 5, the failure prediction candidate list generation unit 802 determines that there is no region in which the vehicle type A is predicted to have a failure in the future.

  In the present embodiment, the bidirectional cluster dividing device 100 uses multivariate data as data indicating which vehicle type in which region has a failure, multivariate data, and history of failure components as sequence data. In contrast, bidirectional clustering is performed. By performing bi-directional clustering on multivariate data and sequence data, it is possible to divide into clusters with common characteristics in vehicle type, region, and fault component history. Can be found. From the clustering result, it is possible to predict an area where a failure is predicted to occur in the future for each vehicle type. The server system 800 can prepare for the occurrence of a failure by transmitting information indicating what type of failure is likely to occur in which vehicle type to a client system in a region where the occurrence of the failure is predicted. In addition, an investigation for finding the cause of the failure can be performed at an early stage.

  Here, in bidirectional clustering, it is usually necessary to set the number of clusters in advance. In the present embodiment, the number of clusters represents the total number of failures occurring. When the number of clusters is set in advance by bidirectional clustering, the number of clusters must be determined in advance even if it is unknown how many failures have occurred as a whole. In other words, by performing clustering, it is necessary to determine in advance the number of faults that have occurred, even though it is desired to know the number of faults that have occurred. In the present embodiment, since the bidirectional cluster dividing apparatus 100 has the cluster number calculation means 103 (FIG. 1), cluster division is performed with an appropriate number of divisions even if the number of clusters is not determined in advance. Can do. In application, in bi-directional clustering, it is important to output a clustering result divided into clusters by an appropriate number just by inputting data.

  In each of the above embodiments, the number of variables of multivariate data is two, but the number of variables is not limited to two. Further, the combination of multivariate data and sequence data is not limited to that used in the above embodiments. For example, “customer” and “product” may be used as the variables of the multivariate data, and “product purchase history” may be used as the sequence data. Or, use “customer” and “company name” as variables of multivariate data, use “change of job history” as sequence data, and use “product” and “Web page” as variables of multivariate data. As “history of the date and time when each product was introduced and introduced on the web page” may be used. Furthermore, it is possible to use “parts” and “part manufacturers” as variables of multivariate data, and “history of parts delivered by parts manufacturers” can be used as sequence data. “Internet virus name” and “region where Internet virus infection was confirmed” can be used as the sequence data, and “the history of the number of virus infections reported per day” can be used as the sequence data.

  In FIG. 1, the bidirectional cluster dividing device 100 includes the cluster number calculating unit 103, but a configuration without the cluster number calculating unit 103 is also possible. In this case, the bidirectional clustering means 102 may perform cluster division with a preset number of cluster divisions. Further, the bidirectional clustering means 102 and the cluster number calculating means 103 do not have to be provided in the same apparatus, and the bidirectional clustering means 102 and the cluster number calculating means 103 are divided into separate apparatuses, and clustering is performed. And the evaluation of the clustering result may be performed by different apparatuses.

  In each of the above embodiments, the example in which multivariate data and sequence data are input from the outside to the bidirectional cluster dividing device 100 has been described. However, the generation of multivariate data and sequence data is performed in the bidirectional cluster dividing device 100. You may go on. For example, in the second embodiment, the clustering control unit 307 of the Web server 300 (FIG. 11) does not generate multivariate data and sequence data, but stores the request history of each user read from the user request storage unit 302. , And output to the bidirectional cluster dividing device 100 via the output device 308. The bidirectional cluster dividing device 100 is provided with data generation means. Based on the information input from the clustering control unit 307, the bidirectional cluster dividing device 100 generates multivariate data indicating which user has clicked which web advertisement and content visit history until the web advertisement is clicked. Thereafter, bidirectional clustering may be performed.

  As described above, the present invention has been described based on the preferred embodiment. However, the bidirectional cluster dividing device, the advertisement distribution system, the product recommendation system, the failure prediction system, the method, and the program of the present invention are limited to the above embodiment. The present invention is not limited, and modifications and changes made from the configuration of the above embodiment are also included in the scope of the present invention.

  Finally, the outline of the present invention will be described. FIG. 29 shows an outline of the bidirectional cluster dividing apparatus of the present invention. The bidirectional cluster dividing device 10 includes an input unit 11 and a bidirectional clustering unit 12. The input means 11 inputs variable data and sequence data corresponding to multivariate data. The bidirectional clustering means 12 performs bidirectional clustering on multivariate data and sequence data. The bi-directional clustering means 12 converts the multivariate data and the sequence data into a plurality of clusters using an evaluation function that indicates whether there are many or less common features between the variables included in the cluster and between the sequence data. Divide into

  In the present invention, not only multivariate data but also sequence data corresponding to the multivariate data is simultaneously subjected to bidirectional clustering. Therefore, it is possible to simultaneously divide into clusters having common features between the variables and between the sequence data. Further, data characteristics appear not only in multivariate data but also in sequence data corresponding to the multivariate data. For this reason, it can be expected that features between multivariate data can be extracted more accurately by simultaneously handling multivariate data and sequence data and performing bidirectional clustering.

10: Bidirectional cluster dividing device 11: Input means 12: Bidirectional clustering means 100: Bidirectional cluster dividing device 101: Input means 102: Bidirectional clustering means 103: Cluster number calculating means 104: Output means 200: User terminal 201: Content request unit 202: Content display unit 300: Web server 301: Content distribution unit 302: User request storage unit 303: Content storage unit 304: Advertisement selection unit 305: Advertisement storage unit 306: Request reception unit 307: Clustering control unit 308: Output device 309: Input device 310: Clustering result storage unit 400: Internet 401, 402: Networks 501-503, 701-703: Client system 600: Server system 601: Data raw Means 602: Recommendation Product list generating means 603: clustering result storage unit 800: the server system 801: data generating means 802: failure prediction candidate list generating unit 803: clustering result storage unit

Claims (21)

Input means for inputting multivariate data and sequence data corresponding to the multivariate data;
Bidirectional clustering is performed on the multivariate data and the sequence data, and the multivariate data and the sequence data have many or less features common to each variable included in the cluster and between the sequence data. A bi-directional cluster dividing device comprising bi-directional clustering means for dividing into a plurality of clusters using an evaluation function representing   The evaluation function is a function whose value decreases as the number of common features between each variable included in the cluster and between the sequence data increases, and the bidirectional clustering means calculates the value of the evaluation function for each cluster. The bidirectional cluster dividing apparatus according to claim 1, wherein cluster division is performed so that a sum total of evaluation function values for each cluster is small.   The cluster number calculating means for determining the number of cluster divisions of bidirectional clustering performed by the bidirectional clustering means based on the value of the evaluation function after the bidirectional clustering means performs cluster division. 3. The bidirectional cluster dividing device according to 2.   4. The bidirectional cluster dividing apparatus according to claim 3, wherein the bidirectional clustering means performs cluster division with the cluster division number determined by the cluster number calculating means when the cluster number calculating means increases the cluster division number. A request accepting means for accepting a request for content from a user, and storing the requested user and the requested content in a user request storage unit;
A content distribution means for transmitting the content requested by the user with an advertisement including a mechanism for allowing the user to request the content of the advertiser;
Based on the information stored in the user request storage unit, the user and the advertisement are used as variables, and multivariate data indicating whether or not the user has requested the content of the advertiser from the advertisement is generated. Correspondingly, a data generation means for generating sequence data in which the requests sent until the user requests the advertiser's content are arranged in time series,
Is bi-directional clustering performed on the multivariate data and the sequence data, and whether the multivariate data and the sequence data have many common features between each variable and sequence data included in the cluster? A bi-directional clustering means for dividing into a plurality of clusters using an evaluation function representing a small number and outputting bi-directional clustering results;
An advertisement distribution system comprising: an advertisement selection unit that determines an advertisement to be added to the content by the content distribution unit based on the bidirectional clustering result. Collect sales information including information that the customer has purchased the product, and based on the collected sales information, variable the customer and the product, and generate multivariate data indicating whether the customer has purchased the product In addition, in response to the multivariate data, data generation means for generating sequence data in which histories related to the purchase of a product by a customer are arranged in time series, and
Bidirectional clustering is performed on the multivariate data and the sequence data, and the multivariate data and the sequence data have many or less features common to each variable included in the cluster and between the sequence data. Bi-directional clustering means for dividing the multi-cluster into a plurality of clusters using an evaluation function representing
A product recommendation system comprising recommended product list generation means for determining a product recommended for a customer based on the bidirectional clustering result. Multivariate that collects failure information including vehicle type and failure location of the vehicle, and based on the collected failure information, makes the vehicle type and region variable, and indicates whether or not a failure has occurred in the region for the vehicle type Data generating means for generating data, and corresponding to the multivariate data, generating data of sequence data in which the history of failure locations that occurred in the past in the vehicle type is arranged in time series,
Bidirectional clustering is performed on the multivariate data and the sequence data, and the multivariate data and the sequence data have many or less features common to each variable included in the cluster and between the sequence data. Bi-directional clustering means for dividing the multi-cluster into a plurality of clusters using an evaluation function representing
A failure prediction system comprising failure prediction candidate list generation means for estimating a region where a failure is predicted to occur for a vehicle type based on the bidirectional clustering result. Inputting multivariate data and sequence data corresponding to the multivariate data to a computer;
The computer performs bi-directional clustering on the multivariate data and the sequence data, and whether the multivariate data has many or few features common to each variable included in the cluster and between the sequence data. A bi-directional cluster partitioning method comprising the step of partitioning into a plurality of clusters using a representing evaluation function   The evaluation function is a function whose value decreases as the number of features common to each variable included in the cluster and between the sequence data increases, and the computer performs evaluation for each cluster in the step of performing the bidirectional clustering. 9. The bidirectional cluster dividing method according to claim 8, wherein the function is calculated and the cluster is divided so that the sum of evaluation function values for each cluster becomes small.   10. The bidirectional cluster according to claim 8, further comprising a step of determining a number of cluster divisions for bidirectional clustering based on a value of the evaluation function after the step of performing the bidirectional clustering. Split method.   The computer according to claim 10, wherein in the step of determining the number of cluster divisions, when it is determined that the number of cluster divisions is increased from the current number of cluster divisions, bi-directional clustering is further performed using the determined number of cluster divisions. Bidirectional cluster partitioning method. A request receiving step in which a computer receives a request for content from a user, and stores the requested user and the requested content in a user request storage unit;
A content distribution step in which a computer adds an advertisement including a mechanism for causing the user to request the advertiser's content to the content requested by the user;
Based on the information stored in the user request storage unit, the computer uses the user and the advertisement as variables, and generates multivariate data indicating whether the user has requested the advertiser's content from the advertisement. In response to multivariate data, a data generation step for generating sequence data in which the requests sent before the user requests the content of the advertiser are arranged in time series,
Whether the computer performs bi-directional clustering on the multivariate data and the sequence data, and the multivariate data has many or less common features between the variables included in the cluster and between the sequence data A bi-directional clustering step that divides the data into a plurality of clusters using an evaluation function that represents and outputs bi-directional clustering results;
An advertisement distribution method comprising: an advertisement selection step in which a computer determines an advertisement to be added to the content based on the bidirectional clustering result. Multivariate data in which a computer collects sales information including information that a customer has purchased a product, and based on the collected sales information, the customer and the product are variables, and the customer has purchased the product. And generating data corresponding to the multivariate data and generating sequence data in which histories relating to the purchase of the product by the customer are arranged in time series, and
The computer performs bi-directional clustering on the multivariate data and the sequence data, and the multivariate data and the sequence data are common to each variable included in the cluster and between the sequence data. A bi-directional clustering step that divides the data into a plurality of clusters using an evaluation function that indicates whether there are many or few, and outputs a bi-directional clustering result;
A product recommendation method comprising: a recommended product list generation step in which a computer determines a product recommended for a customer based on the bidirectional clustering result. The computer collects failure information including the vehicle type and the failure location of the vehicle, and based on the collected failure information, the vehicle type and the region are variables, and whether or not a failure has occurred in the region for the vehicle type. A data generation step of generating sequence data in which the history of failure locations that occurred in the past in the vehicle type is arranged in time series in response to the multivariate data,
The computer performs bi-directional clustering on the multivariate data and the sequence data, and the multivariate data and the sequence data are common to each variable included in the cluster and between the sequence data. A bi-directional clustering step that divides the data into a plurality of clusters using an evaluation function that indicates whether there are many or few, and outputs a bi-directional clustering result;
A failure prediction method comprising: a failure prediction candidate list generation step in which a computer estimates a region where a failure is predicted to occur for a vehicle type based on the bidirectional clustering result. On the computer,
Processing to input multivariate data and sequence data corresponding to the multivariate data;
An evaluation function that performs bidirectional clustering on the multivariate data and the sequence data, and indicates whether the multivariate data has many or few features common to each variable included in the cluster and between the sequence data A program that executes processing to divide into a plurality of clusters using.   The evaluation function is a function having a value that decreases as the number of features common to each variable and sequence data included in the cluster increases. In the process of performing bidirectional clustering, the value of the evaluation function is determined for each cluster. The program according to claim 15, wherein cluster division is performed so that the sum of evaluation function values for each cluster is reduced.   The program according to claim 15 or 16, further causing the computer to further execute a process of determining a cluster division number of the bidirectional clustering based on the value of the evaluation function after the process of performing the bidirectional clustering.   The program according to claim 17, wherein in the process of determining the number of cluster divisions, if it is determined that the number of cluster divisions is to be increased from the current number of cluster divisions, bi-directional clustering is further executed with the determined number of cluster divisions. On the computer,
A request reception process for receiving a request for content from a user, and storing the request transmission user and the requested content in a user request storage unit;
A content distribution process in which an advertisement including a mechanism for allowing the user to request the content of the advertiser is added to the content requested by the user and transmitted;
Based on the information stored in the user request storage unit, the user and the advertisement are used as variables, and multivariate data indicating whether or not the user has requested the content of the advertiser from the advertisement is generated. Correspondingly, a data generation process for generating sequence data in which the requests sent until the user requests the advertiser's content are arranged in time series,
Bidirectional clustering is performed on the multivariate data and the sequence data, and the multivariate data is evaluated to indicate whether there are many or few features common to each variable included in the cluster and between the sequence data. Bidirectional clustering processing that divides into multiple clusters using functions and outputs bidirectional clustering results;
A program for executing an advertisement selection process for determining an advertisement to be added to the content based on the bidirectional clustering result. On the computer,
Collect sales information including information that the customer has purchased the product, and based on the collected sales information, variable the customer and the product, and generate multivariate data indicating whether the customer has purchased the product In addition, in response to the multivariate data, a data generation process for generating sequence data in which histories relating to the purchase of a product by a customer are arranged in time series, and
Bidirectional clustering is performed on the multivariate data and the sequence data, and the multivariate data and the sequence data have many or less features common to each variable included in the cluster and between the sequence data. A bi-directional clustering process that divides the data into a plurality of clusters using an evaluation function that expresses
A program for executing a recommended product list generation process for determining a product recommended for a customer based on the bidirectional clustering result. On the computer,
Multivariate that collects failure information including vehicle type and failure location of the vehicle, and based on the collected failure information, makes the vehicle type and region variable, and indicates whether or not a failure has occurred in the region for the vehicle type A data generation process for generating data, and corresponding to the multivariate data, generating sequence data in which histories of fault locations that occurred in the past in the vehicle type are arranged in time series,
Bidirectional clustering is performed on the multivariate data and the sequence data, and the multivariate data and the sequence data have many or less features common to each variable included in the cluster and between the sequence data. A bi-directional clustering process that divides the data into a plurality of clusters using an evaluation function that expresses
A program that executes a failure prediction candidate list generation process that estimates a region where a failure is predicted to occur for a vehicle type based on the bidirectional clustering result.

Images