JP2016173784A - Tensor factorization processing device, tensor factorization processing method, and tensor factorization processing program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To dynamically determine a correct processing procedure for any input condition in tensor factorization processing.SOLUTION: A tensor factorization processing device 1 for performing tensor factorization processing includes: a tensor construction unit 11 for constructing a tensor having, as mode, attribute information extracted from a group of log data including a plurality of pieces of attribute information of a specific event; an auxiliary data construction unit 12 for constructing auxiliary data formed of tensor information including a factor matrix of each tensor and factor matrix information including a sequence in the tensor for each factor matrix; and a reference determination unit 21 for determining a factor matrix to be referred to during the update of the factor matrix on the basis of the auxiliary data.SELECTED DRAWING: Figure 1

Description

  The present invention relates to a data mining technique, and more particularly to a technique related to factorization for extracting a factor pattern from a plurality of attribute information.

  As a technique for extracting a factor pattern from a plurality of attribute information, there is a technique called non-negative composite tensor factorization (for example, Non-Patent Document 1). This technology is formulated as a mathematical formula. If the input conditions are determined, a factorization processing procedure corresponding to the input conditions is set by using software for numerical calculation (factor matrix). ) Can be obtained.

Koh Takeuchi, Ryota Tomioka, Katsuhiko Ishiguro, Akisato Kimura, and Hiroshi Sawada, "Non-negative Multiple Tensor Factorization", ICDM, 2013

  In the prior art, there is no method for dynamically determining the factorization processing procedure of a tensor as a mathematical expression for an arbitrary input condition, and the processing procedure must be set each time according to the input condition.

  As the difference in the input conditions and the influence on the processing procedure, for example, the number of factor matrices to be output differs depending on the number of input tensors and ranks. In addition, when a plurality of tensors have a shared mode, the values and the number of factor matrices of the tensors to be referred to when the tensor is factorized and updated are different.

  In view of the above circumstances, an object of the present invention is to dynamically determine a correct processing procedure for arbitrary input conditions in tensor factorization processing.

  Therefore, the present invention performs a tensor factorization process based on a tensor constructed from log data including attribute information of a specific event and its auxiliary data.

  An aspect of the present invention as a tensor factorization processing apparatus is a tensor factorization processing apparatus that performs tensor factorization processing, and uses attribute information extracted from a group of log data including attribute information of a specific event as a mode. Tensor construction means for constructing a tensor, and auxiliary data construction means for constructing auxiliary data composed of tensor information including the factor matrix of the tensor for each tensor and factor matrix information including the order in the tensor for each factor matrix And reference determination means for determining a factor matrix to be referred to when updating the factor matrix based on the auxiliary data.

  An aspect as a tensor factorization processing method of the present invention is a tensor factorization processing method executed by a tensor factorization processing apparatus that performs tensor factorization processing, and includes a log data group including attribute information of a specific event. Auxiliary steps consisting of a tensor construction step that constructs a tensor with the extracted attribute information as a mode, tensor information including the factor matrix of the tensor for each tensor, and factor matrix information including the order in the tensor for each factor matrix An auxiliary data construction step of constructing data; and a reference determination step of determining a factor matrix to be referred to when updating the factor matrix based on the auxiliary data.

  Note that the present invention may be in the form of a program that causes a computer to function as each unit of the apparatus or a program that causes a computer to execute each step of the method.

  According to the above invention, the correct processing procedure for an arbitrary input condition can be dynamically determined in the factorization processing of the tensor.

(A) is a block block diagram of the tensor factor decomposition processing apparatus in the embodiment of the present invention, (b) is a block block diagram of a tensor decomposition unit in the apparatus. The data block diagram which showed an example of log data. The data block diagram which showed an example of the tensor based on log data. Explanatory drawing of the process in which a tensor is constructed from log data. (A) is a data block diagram which showed an example of the tensor information of auxiliary data, (b) is a data block diagram which showed an example of factor row example information of the auxiliary data. The flowchart explaining the process of building a tensor. The flowchart explaining the process of constructing auxiliary data. The flowchart explaining the process which initializes the factor matrix of a tensor. The flowchart explaining the determination process of the factor matrix which should be referred. The flowchart explaining the process which updates the said factor matrix.

  Hereinafter, embodiments of the present invention will be described with reference to the drawings. However, the present invention is not limited to these embodiments.

[Overview]
The tensor factorization processing apparatus 1 of the present embodiment shown in FIG. 1 constructs a data structure for determining a tensor factor matrix update processing procedure. The correct update process procedure of the tensor factor matrix for an arbitrary input condition is dynamically determined by referring to this data structure in a certain procedure.

[Explanation of technical terms]
In describing the present embodiment, technical terms related to the present embodiment will be described.

  The log data is data including a combination of a plurality of attribute information of a specific event and attribute values corresponding to the combination. For example, log data indicating that people have visited a store includes three attributes of a user ID, a store ID, and a day of the week, and attribute values indicated by the number of visits and the stay time corresponding to these attributes. The log data can be expressed as a tensor by regarding the attribute as a mode.

  A tensor is synonymous with a multidimensional array in the present invention. For example, the third-floor tensor can be expressed as a three-dimensional array.

  Mode refers to the tensor axis. For example, a matrix can be regarded as a tensor of the second floor, but at this time, there are two modes of a row direction and a column direction. When log data is represented by a tensor, each mode has a one-to-one correspondence with the attribute of the log data.

  The shared mode is a mode that appears in a plurality of tensors among the input tensors. For example, when there is a tensor having a mode of (user ID, store ID, day of the week) and a tensor having a mode of (user ID, environment type, time zone) as an input, the user ID becomes the sharing mode.

  The factor matrix is a matrix obtained by factorizing the tensor, and there are as many as the modes. In the above input example, five factor matrices are obtained corresponding to the five modes.

[Device configuration]
The tensor factorization processing apparatus 1 includes an input data storage unit 10, a tensor construction unit 11, an auxiliary data construction unit 12, a tensor decomposition unit 13, and an output data storage unit 14.

  The input data storage unit 10 stores log data, which is original data of a tensor to be factorized, and parameters used in non-negative composite tensor factorization. Note that log data is stored in advance.

  An example of log data stored in the input data storage unit 10 is shown in FIG. The log data of this example includes a combination of attribute information such as a user ID and a store ID and attribute values corresponding to the combination. However, when a format that includes only attribute information and does not have an attribute value is also assumed, duplication of the same combination in the log data is counted, and the count value is used as the attribute value.

  The tensor construction unit 11 extracts a group of log data from the input data storage unit 10 and constructs a tensor using the attribute information of a specific event extracted from the log data group as a mode. An example of the constructed tensor is shown in FIGS. Detailed processing will be described later.

  The auxiliary data construction unit 12 constructs auxiliary data based on the tensor constructed by the tensor construction unit 11. This auxiliary data consists of tensor information and factor matrix information. The tensor information includes a weight associated with each tensor and a factor matrix of the tensor. The factor matrix information includes an in-tensor order for each factor matrix. An example of the constructed auxiliary data is shown in FIG. Detailed processing will be described later.

  The tensor decomposition unit 13 performs tensor factorization using the tensor and the auxiliary data. As shown in FIG. 1, the tensor decomposition unit 13 includes an initialization unit 20, a reference determination unit 21, a matrix update unit 22, and a calculation end evaluation unit 23 as specific components.

  The initialization unit 20 initializes the factor matrix of the tensor constructed by the tensor construction unit 11 with a random number. Detailed processing will be described later.

  The reference determination unit 21 determines a factor matrix to be referred to when updating the factor matrix of the initialized tensor based on the auxiliary data constructed by the auxiliary data construction unit 12. Detailed processing will be described later.

  The matrix update unit 22 updates the elements of the factor matrix to be referred to that are determined by the reference determination unit 21. Detailed processing will be described later.

  The calculation end evaluation unit 23 determines whether to continue or end the update based on a predetermined criterion.

  The output data storage unit 14 stores the factor matrix obtained by the tensor decomposition unit 13.

  The functional units 10 to 14 and 20 to 23 of the tensor factorization processing apparatus 1 described above are realized by hardware resources of a computer. That is, the tensor factorization processing apparatus 1 includes hardware resources related to a computer such as at least an arithmetic device (CPU), a storage device (memory, a hard disk device, etc.), a communication interface, and the like. And these hardware resources cooperate with software resources (OS, application, etc.), and each function part 10-14, 20-23 is mounted. Moreover, you may make it each implement | achieve the function parts 10-14 and 20-23 in each computer.

[Process of tensor factorization processing of this embodiment]
This process includes the following processes of “tensor construction”, “construction of auxiliary data”, “initialization of factor matrix”, “determination of factor matrix to be referred to”, and “update of factor matrix”.

(Tensor construction)
The steps S100 to S106 of tensor construction will be described with reference to FIG. This process is executed by the tensor construction unit 11.

  S100: A tensor whose element values are all 0 is prepared for each log data extracted from the input data storage unit 10, and the processes after S101 are executed. When the process is completed for all log data, the process proceeds to the end.

  S101: The processing from S102 onward is executed for each row of one log data. When the processing for all the rows is completed, the process proceeds to S100.

  S102: The processing from S103 onward is executed for each attribute of one line. When the processing is completed for all attributes, the process proceeds to S106.

  S103: If an attribute value of a certain attribute has been assigned to a subscript in S105 in the past, the process proceeds to S104, and if not, the process proceeds to S105.

  S104: The attribute value is converted into a subscript assigned in the past, and the process proceeds to S102.

  S105: An unassigned subscript is assigned to the attribute value, the attribute value is converted into the subscript, and the process proceeds to S102.

  S106: The attribute value is assigned to the element indicated by the subscript converted in S103 to S106 of the tensor.

  The tensors T0 and T1 illustrated in FIG. 3 and the tensor T0 illustrated in FIG. 4 constructed in the process of S100 to S106 described above are associated with the weight values as illustrated in FIG. It is stored in the storage unit 10.

(Construction of auxiliary data)
The auxiliary data construction steps S200 to S207 will be described with reference to FIG. This process is executed by the auxiliary data construction unit 12.

  S200: The processing after S201 is executed for each tensor stored in the input data storage unit 10. When the processing is completed for all tensors, the processing is terminated.

  S201: A tensor is extracted from the input data storage unit 10 together with its weight value.

  S202: The log data attribute name (in the example of FIG. 4, user ID, store ID, day of the week, environment type, time zone) corresponding to each mode of the tensor is extracted from the input data storage unit 10.

  S203: The processing after S204 is executed for each mode of the tensor. When processing is completed for all modes of all tensors stored in the input data storage unit 10, the process proceeds to S207.

  S204: If there is factor matrix information having the same attribute name for a certain mode, the process proceeds to S205, and if there is no factor matrix information, the process proceeds to S206.

  S205: The tensor information number and the in-tensor order of the factor matrix of the tensor corresponding to this number are added to the factor matrix information. The total number of tensor information created at this point is used as the tensor information number. For example, it is 0 when no tensor information has been created yet.

  S206: New factor matrix information is added, and the attribute name of the mode, the tensor number having the mode as a component, and the in-tensor order of the factor matrix of the tensor corresponding to this number are added to this factor matrix information as in S205. .

  S207: New tensor information is added, the weight read in S201 is associated with this tensor information, and the number of the factor matrix information selected or added in S205 and S206 is added.

  The auxiliary data constructed in the process of S200 to S207 described above is, for example, an input data storage unit 10 in a form composed of tensor information shown in FIG. 5A and factor matrix information shown in FIG. Saved in.

(Factor matrix initialization)
The factor matrix initialization steps S300 to S302 will be described with reference to FIG. This process is executed by the initialization unit 20 of the tensor decomposition unit 13.

  S300: The rank number (rank) is extracted from the input data storage unit 10 as a parameter used in the factorization of this embodiment.

  S301: The processing of S302 is executed for each factor matrix information of the auxiliary data constructed by the auxiliary data construction unit 12. When the process is completed for all factor matrix information, the process proceeds to the end.

  S302: Random numbers greater than 0 are substituted for all elements of the factor matrix of the tensor corresponding to each factor matrix information. In the factor matrix, the row size is the size of the corresponding mode, and the column size is the rank number derived in S300.

(Determining the factor matrix to be referenced)
The process S400 to S404 for determining a factor matrix to be referred to will be described with reference to FIG. This process is executed by the reference determination unit 21 of the tensor decomposition unit 13.

  S400: The processing after S401 is executed for each tensor information of the auxiliary data in which the factor matrix is initialized. When the process is completed for all tensor information, the process proceeds to the end.

  S401: Processes after S402 are executed for each factor matrix information corresponding to each tensor information. When the processing is completed for all factor matrix information, the process proceeds to S400.

  S402: For each tensor information of the factor matrix information, an estimated value of the corresponding tensor is calculated. The estimated value of the tensor corresponding to the tensor information can be calculated from the factor matrix corresponding to the number of the factor matrix information of the tensor information. In the example of FIG. 5A, the estimated value of tensor 0 is calculated based on the values of the elements of the factor matrix corresponding to the factor matrix numbers 0, 1, and 2. The mathematical definition of this estimated value is disclosed in Non-Patent Document 1.

  S403: For each tensor information of the factor matrix information, select factor matrix information having an order other than the corresponding intra-tensor order. In the case of FIG. 5, when the factor matrix information 1 is processed, the factor matrix information corresponding to the orders 0 and 2 in which the order in the tensor 0 is other than 1 for the tensor information 0, that is, the factor matrix information 0 and 2 is selected. . Similarly, when processing is performed on the factor matrix information 0, the factor matrix information 1 and 2 corresponding to the orders 1 and 2 in which the order in the tensor 0 is other than 0 for the tensor information 0, and the order in the tensor 1 is other than 0 for the tensor information 1 Factor matrix information 3 and 4 corresponding to orders 1 and 2 are selected.

  S404: The matrix updating unit 22 is designated with one or more sets of factor matrix information numbers to be updated and (a list of tensor information numbers and factor matrix information numbers). In the case of FIG. 5, when the factor matrix information 0 is processed, “0, (0, 1, 2), (1, 3, 4)” is designated. When the factor matrix information 1 is processed, “1, (0, 0, 2)” is designated.

(Update factor matrix)
The process of updating the factor matrix will be described with reference to FIG. This process is executed by the matrix update unit 22 of the tensor decomposition unit 13.

  S500: The processing from S501 is performed on each element of the factor matrix corresponding to the designated factor matrix information. When the process is completed for all elements, the process proceeds to the end.

  S501: Based on the list of the specified tensor information number and factor matrix information number, the value of the update formula of the factor matrix element is calculated, and the element is updated. That is, the value of the update formula of the element matrix element of the tensor is calculated in the order of the number of the factor matrix information corresponding to the tensor of the designated tensor information number, and the element is updated.

  The update formula varies depending on the definition of the distance between the tensors. For example, when the generalized KL divergence is the distance, the update formula for the elements of the factor matrix related to the two tensors is expressed by the following formula (1). The parameters necessary for the calculation of the update formula are the value of the tensor weight η corresponding to the tensor information, the estimated value of the tensor calculated in S402, and the factor matrix of the tensor obtained in S403.

  Then, the calculation end evaluation unit 23 determines whether to end or continue the update (S500, S501) based on the factor matrix updated in S501.

  Specifically, the calculation end evaluation unit 23 determines, for each tensor, the distance between the “estimated value of the tensor that can be calculated from the updated factor matrix” and the “first tensor constructed by the tensor construction unit 11” or the number of times of the update. Whether to continue or end the update is determined based on the above. For example, a KL divergence distance is adopted as the distance. This generalized KL divergence is an example, and other stochastic known distances may be employed. If another distance is adopted here, Equation (1) is also suitable for that distance.

  Then, when the distance does not satisfy the termination condition, or when the number of times does not reach the upper limit, it is determined that the update is continued.

  On the other hand, when the distance satisfies a preset end condition, or when the number of times reaches a preset upper limit, the end of the update is determined.

  The tensor factor matrix updated as described above is stored in the output data storage unit 14.

[Effect of this embodiment]
According to the tensor factorization processing apparatus 1 of the present embodiment, by referring to a data structure for determining a tensor factor matrix update process procedure in a certain procedure, a correct update process of a factor matrix for an arbitrary input condition is performed. The procedure can be determined dynamically. Thereby, it is not necessary to change the code to cope with different input conditions, and data mining for various conditions can be performed efficiently.

  Therefore, for example, when the number of tensor factor matrices to be output differs depending on the number of input tensors and ranks, or when performing update processing by factorizing multiple tensors having shared modes, Even when the values of tensor factor matrices and their numbers are different, the correct update procedure for factorization can be determined dynamically.

  The tensor factorization process of the above embodiment is an example of a third-order tensor, but the processing procedure is determined in the same manner even when a part of the input is a matrix or a tensor of the fourth or higher floor. Can do. This makes it possible to handle various data such as big data.

  Further, when the factor matrix of the auxiliary data is updated, in step S403, when updating the factor matrix of the auxiliary data, a factor matrix corresponding to an order other than the order in the tensor of the factor matrix is referred to as a factor matrix to be referred to. decide. Thereby, the factor matrix to be updated in the update can be dynamically determined.

  In the process of constructing the auxiliary data (S200 to S207), the tensor is stored in the tensor information in association with the weight. Thereby, the element of the factor matrix reflecting the weight of the tensor of the factor matrix can be updated.

[Other Embodiments of the Present Invention]
The present invention can be realized by configuring a program that causes a computer to function as a part or all of the functional units 11 to 13 and 20 to 23 constituting the tensor factorization processing apparatus 1, and causing the computer to execute the program. Alternatively, a part or all of the above steps S100 to S106, S200 to S207, S300 to S302, S400 to S404, S500, and S501 executed by the apparatus 1 are configured and executed by the computer. This can be realized. The program can be provided by being stored in a known recording medium (for example, a hard disk, a flexible disk, a CD-ROM, etc.) that can be read by the computer. Alternatively, the program can be provided via the network via the Internet or e-mail.

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

DESCRIPTION OF SYMBOLS 1 ... Tensor factorization processing apparatus 11 ... Tensor construction part (tensor construction means)
12 ... Auxiliary data construction unit (Auxiliary data construction means)
13 ... tensor decomposition unit 20 ... initialization unit 21 ... reference determination unit (reference determination means)
22 ... Matrix update unit 23 ... Calculation end evaluation unit

Claims (7)

A tensor factorization processing apparatus that performs tensor factorization processing,
A tensor construction means for constructing a tensor having a mode of attribute information extracted from a group of log data including a plurality of attribute information of a specific event;
Auxiliary data construction means for constructing auxiliary data composed of tensor information including the factor matrix of the tensor for each tensor and factor matrix information including the order in the tensor for each factor matrix;
A tensor factorization processing apparatus comprising: a reference determination unit that determines a factor matrix to be referred to when updating the factor matrix based on the auxiliary data. 2. The reference determining unit, when updating a factor matrix of the auxiliary data, determines a factor matrix corresponding to an order other than the order in the tensor of the factor matrix as the factor matrix to be referred to. The tensor factorization processing apparatus according to 1.   The tensor factorization processing apparatus according to claim 1, wherein the auxiliary data construction unit stores the tensor in the tensor information in association with the weight. A tensor factorization processing method executed by a tensor factorization processing apparatus that performs tensor factorization processing,
A tensor construction step for constructing a tensor having a mode of attribute information extracted from a group of log data including a plurality of attribute information of a specific event;
Auxiliary data construction step of constructing auxiliary data consisting of tensor information including the factor matrix of the tensor for each tensor and factor matrix information including the order in the tensor for each factor matrix;
A tensor factorization processing method, comprising: a reference determination step of determining a factor matrix to be referred to when updating the factor matrix based on the auxiliary data. 5. In the reference determination step, when updating the factor matrix of the auxiliary data, a factor matrix corresponding to an order other than the order in the tensor of the factor matrix is determined as the factor matrix to be referred to. The tensor factorization processing method described in 1. 6. The tensor factorization processing method according to claim 4, wherein, in the auxiliary data construction step, the tensor is stored in the tensor information in association with the weight.   A tensor factor decomposition processing program for causing a computer to function as each means constituting the tensor factor decomposition processing device according to any one of claims 1 to 3.

Images