Feature creation device, feature creation method, and program

Abstract

A feature quantity creation device (10) according to the present disclosure is provided with: an interface (11) for acquiring data of an object of prediction and a first feature quantity regarding the object of prediction; a weight matrix generation unit (13) that generates a weight matrix having a number of rows equal to the number of columns of a feature quantity matrix in which a first feature quantity for each row element as generated on the basis of said acquired first feature quantity is arranged in the column direction, and having a number of columns that is greater than the maximum number of column elements assumed for each row element; and a feature quantity creation unit (14) that creates a second feature quantity on the basis of a statistical feature quantity matrix that is the product of the feature quantity matrix and the weight matrix.

Description

[Technical Field]

[0001] The present disclosure relates to a feature creation device, a feature creation method, and a program. [Background technology]

[0002] Deep learning predictions are being adopted in a wide range of fields, such as demand forecasting for product sales and road congestion forecasting. For example, in the field of communications, it is possible to design an efficient communications network by estimating the bandwidth of links required for the network based on predictions of future communications traffic. Non-Patent Document 1 describes a technology for highly accurate traffic prediction using deep learning. In the technology described in Non-Patent Document 1, a model is trained based on a feature vector consisting of contract information for each line, and the model is optimized based on the total traffic of multiple lines. Then, in the technology described in Non-Patent Document 1, the total traffic of multiple lines is predicted using a model based on the future contract information for each line.

[0003] In machine learning and deep learning, feature engineering is performed to improve prediction accuracy. A common method of feature engineering is statistical processing of features related to the prediction target. For example, Non-Patent Document 1 describes a technique for performing statistical processing on the series, total value, and average value of the contract information for each line. This technique is based on the premise that there is a correlation between the magnitude of a feature, such as contract bandwidth, and the amount of traffic (the prediction target).

[0004] As a technique for creating features that can be applied even when the correlation between the prediction target and features is low, Non-Patent Document 2 describes a technique for combining numerically converted features. In this technique, statistical features suitable for prediction are created by a conversion process in which weighted features are added. [Prior art documents] [Non-patent literature]

[0005]

Non-Patent Document 1

Non-Patent Document 2

Summary of the Invention

Problems to be Solved by the Invention

[0006] In the statistical process of searching for the optimal numerical conversion of feature quantities as described in Non-Patent Document 2, it is necessary to optimize the feature quantities while repeating the creation of feature quantities by numerical conversion, the learning of a model based on the feature quantities, and the testing of the model, which is a time-consuming problem. Also, in order to search for what kind of numerical conversion to use, the skills of engineers are required.

[0007] In view of the above problems, an object of the present disclosure is to provide a feature quantity creation device, a feature quantity creation method, and a program that can more easily create new feature quantities by numerical conversion even when the correlation between the prediction target and the feature quantities is small.

Means for Solving the Problems

[0008] In order to solve the above problem, a feature creation device according to the present disclosure is a feature creation device that is used for training a model using deep learning and creates features related to an object to be predicted by the model, and includes: an acquisition unit that acquires data of the object to be predicted and first features related to the object to be predicted; a weight matrix generation unit that generates a weight matrix whose number of rows is the same as the number of columns of a feature matrix in which the first features for one row element are arranged in the column direction, the feature matrix being generated based on the acquired first features, and whose number of columns is greater than the maximum number of column elements expected for the one row element; and a feature creation unit that creates second features based on a statistical feature matrix that is the product of the feature matrix and the weight matrix.

[0009] Furthermore, in order to solve the above-described problems, a feature creation method according to the present disclosure is a feature creation method by a feature creation device that creates features related to an object to be predicted by a model, the feature creation method being used for training a model using deep learning, and includes the steps of: acquiring data of the object to be predicted and first features related to the object to be predicted; generating a weight matrix whose number of rows is the same as the number of columns of a feature matrix in which the first features for one row element are arranged in the column direction, the weight matrix being generated based on the acquired first features, and whose number of columns is greater than the maximum number of column elements expected for the one row element; and creating second features based on a statistical feature matrix that is the product of the feature matrix and the weight matrix.

[0010] In order to solve the above problem, a program according to the present disclosure causes a computer to operate as the feature creation device described above. [Effects of the Invention]

[0011] According to the feature creation device, feature creation method, and program disclosed herein, even when the correlation between the prediction target and the feature is low, new feature can be created more easily by numerical conversion. [Brief explanation of the drawings]

[0012] [Figure 1]FIG. 1 is a diagram illustrating an example of the configuration of a feature creation device according to an embodiment of the present disclosure. [Figure 2] 2 is a diagram illustrating an example of a feature matrix stored in a storage unit illustrated in FIG. 1. FIG. [Figure 3] 2 is a flowchart showing an example of the operation of the feature creation device shown in FIG. [Figure 4] 2 is a diagram showing an example of a weighting matrix generated by a weighting matrix generation unit shown in FIG. 1. FIG. [Figure 5] 4 is a flowchart showing an example of an operation of generating a weight matrix by a weight matrix generating unit shown in FIG. [Figure 6] 10 is a flowchart showing another example of the operation of generating a weighting matrix by the weighting matrix generation unit shown in FIG. [Figure 7] 2 is a diagram showing an example of a basic weighting matrix generated by a weighting matrix generation unit shown in FIG. 1. FIG. [Figure 8] 2 is a diagram for explaining generation of a weight matrix from a basic weight matrix by a weight matrix generation unit shown in FIG. 1. FIG. [Figure 9] 2 is a diagram for explaining generation of a statistical feature matrix by a statistical feature matrix generation unit shown in FIG. 1. FIG. [Figure 10] FIG. 2 is a diagram illustrating an example of a hardware configuration of the feature creation device illustrated in FIG. DETAILED DESCRIPTION OF THE INVENTION

[0013] Hereinafter, embodiments of the present disclosure will be described with reference to the drawings.

[0014] FIG. 1 is a diagram illustrating an example configuration of a feature creation device 10 according to an embodiment of the present disclosure. The feature creation device 10 according to this embodiment creates features related to a target to be predicted by a model, which are used for training a model using deep learning. The feature creation device 10 according to this embodiment also trains a model using the created features, and predicts a target to be predicted using the trained model. In the following description, the target to be predicted is traffic in a communication network, but the present disclosure is not limited to this example.

[0015] As shown in FIG. 1, a feature creation device 10 according to this embodiment includes an interface 11, a storage unit 12, a weighting matrix generation unit 13, a statistical feature matrix generation unit 14, and a prediction function unit 15.

[0016] The interface 11, which serves as an acquisition unit, acquires data of a prediction target and feature data (first feature) related to the prediction target. In this embodiment, the interface 11 acquires traffic data from a traffic collection device 1 as the data of the prediction target, and acquires feature data from a feature data source 2 as the feature related to the prediction target. The traffic collection device 1 is a device that collects traffic data, which is the sum of traffic flowing through multiple lines between communication devices, at predetermined time intervals. The feature data source 2 is a database in which feature data for each of multiple lines in a communication network is registered. The feature data includes information that affects traffic in the communication network, such as service information indicating services provided for each line and option information indicating options added to each line. This information is registered in the feature data source 2 based on, for example, contract information for each line.

[0017] The interface 11 outputs the acquired traffic data and feature data to the storage unit 12 .

[0018] The storage unit 12 stores traffic data output from the interface 11. The storage unit 12 also stores feature data output from the interface 11 in the form of a feature matrix. Specifically, as shown in FIG. 2, the storage unit 12 stores a feature matrix in which date and time (date in the example shown in FIG. 2) are row elements, multiple lines (three lines (line 1, line 2, and line 3) in the example shown in FIG. 2) are column elements, and feature values ​​(first feature values) for each line for each row element (date) are arranged in the column direction. That is, if the number of date and time (dates) as row elements is N and the number of lines as column elements is M, the storage unit 12 stores the feature data as a feature matrix of N rows by M columns. The storage unit 12 also stores a statistical feature matrix generated by a statistical feature matrix generation unit 14, which will be described later.

[0019] The weight matrix generation unit 13 generates a weight matrix having the same number of rows as the number of columns of a feature matrix in which features (first features) for one row element are arranged in the column direction, the feature matrix being generated based on the acquired feature data (first features), and the number of columns is greater than the maximum number of column elements (number of lines) expected for one row element of the feature matrix. The weight matrix generation unit 13 outputs the feature matrix and the generated weight matrix to the statistical feature matrix generation unit 14. The generation of the weight matrix by the weight matrix generation unit 13 will be described in detail later.

[0020] The statistical feature matrix generation unit 14, which serves as a feature creation unit, generates a statistical feature matrix by multiplying the feature matrix and weight matrix output from the weight matrix generation unit 13, and outputs the generated statistical feature matrix to the storage unit 12. The values ​​of each column of the statistical feature matrix are used as features (second features) for model training by the prediction function unit 15, which will be described later. That is, the statistical feature matrix generation unit 14 generates features (second features) by multiplying the feature matrix and weight matrix. Details of the generation of the statistical feature matrix by the statistical feature matrix generation unit 14 will be described later.

[0021] In the learning phase, the prediction function unit 15 learns a model for predicting a prediction target (traffic) based on the traffic data and statistical feature matrix (second feature) stored in the storage unit 12. In the inference phase, the prediction function unit 15 inputs, into the learned model, feature (second feature) created based on feature data at any timing at which prediction is desired, to predict traffic.

[0022] Next, the operation of the feature creation device 10 according to this embodiment will be described. Fig. 3 is a flowchart showing an example of the operation of the feature creation device 10 according to this embodiment, and is a diagram for explaining a feature creation method by the feature creation device 10.

[0023] The interface 11 acquires traffic data from the traffic collection device 1 and acquires feature data from the feature data source 2 (step S1). The acquired traffic data is stored in the storage unit 12. The acquired feature data is also stored in the storage unit 12 as a feature matrix of N rows (N is the number of dates and times) by M rows (M is the number of lines) as shown in FIG.

[0024] The weight matrix generation unit 13 generates a weight matrix whose number of rows is the same as the number of columns in the feature matrix and whose number of columns is greater than the maximum number of column elements (number of lines) expected for one row element (step S). That is, the weight matrix generation unit 13 generates a weight matrix with M rows by L (L>M) columns, as shown in FIG. 4. As described above, M is the number of lines for one row element (date in the example shown in FIG. 2) in the feature matrix. L is the number of weight patterns, and is, for example, a value greater than Mmax, the maximum number of lines expected in the future. FIG. 4 shows an example of a weight matrix whose number of rows is the same as the number of columns in the feature matrix shown in FIG. 2 (M=3) and whose number of columns is 5 (L=5).

[0025] FIG. 5 is a flowchart showing an example of the operation of generating a weighting matrix by weighting matrix generating section 13. In FIG.

[0026] The weighting matrix generation unit 13 acquires the maximum value Mmax of the number of lines expected in the future (the maximum number of features expected for one item) (step S21). The maximum value Mmax may be stored in advance in the storage unit 12, for example, and the weighting matrix generation unit 13 may acquire the maximum value Mmax stored in the storage unit 12. Alternatively, the maximum value Mmax may be input to the feature creation device 10 via an input device (not shown in FIG. 1) at the start of the processing of step S2, and the weighting matrix generation unit 13 may acquire the input maximum value Mmax.

[0027] Upon acquiring the maximum value Mmax, the weighting matrix generation unit 13 determines L such that L>Mmax (step S22). For example, the weighting matrix generation unit 13 may multiply the maximum value Mmax by a preset value to determine L. Alternatively, the weighting matrix generation unit 13 may determine L to be a value larger than the maximum value Mmax, which is obtained using a random number, for example.

[0028] After determining the value of L, the weighting matrix generation unit 13 generates a weighting matrix of M rows and L columns (step S23). The value of each element of the generated weighting matrix is ​​not yet set.

[0029] Next, the weight matrix generation unit 13 sets a random value within a predetermined range (for example, a range of 0 to 1) to each element of the generated weight matrix of M rows and L columns (step S24).

[0030] Fig. 6 is a flowchart showing another example of the operation of generating a weighting matrix by weighting matrix generation section 13. In Fig. 6, the same processes as those in Fig. 5 are denoted by the same reference numerals, and the description thereof will be omitted.

[0031] After setting the value of L (step S22), the weight matrix generation unit 13 generates a basic weight matrix of L rows and L columns (step S25). The value of each element of the generated basic weight matrix is ​​not yet set.

[0032] Next, the weighting matrix generation unit 13 sets the value of each element of the generated basic weighting matrix (step S26). For example, the weighting matrix generation unit 13 sets the value of the element in the mth row and the lth column according to the following equation. l-m+l(m≧l) l-m+l+L(m <l)

[0033] The processing of steps S25 and S26 generates a basic weight matrix of L rows by L columns as shown in Fig. 7. By setting each element of the basic weight matrix according to the above-mentioned formula, as shown in Fig. 7, the value of each column of the basic weight matrix will not be the same as the value of another column. In other words, one weight pattern will not be the same as another weight pattern. Furthermore, the value of each column of the basic weight matrix will not be a constant multiple of the value of another column. In other words, one weight pattern will not be a constant multiple of another weight pattern.

[0034] 6 again, weighting matrix generation unit 13 uses the first to Mth rows of the generated basic weighting matrix to generate a weighting matrix of M rows x L columns (step S27). Specifically, weighting matrix generation unit 13 extracts the first to Mth rows of the basic weighting matrix and generates a weighting matrix of M rows x L columns, as shown in FIG.

[0035] As described above, the values ​​of each column of the basic weighting matrix are set so that they are not the same as the values ​​of other columns or a constant multiple of the values ​​of other columns. Therefore, even in a weighting matrix generated by cutting out a portion of the basic weighting matrix, the values ​​of one column are not the same as the values ​​of other columns or a constant multiple of the values ​​of other columns. In other words, in the weighting matrix generation method described with reference to FIG. 6, weighting matrix generation unit 13 generates a weighting matrix so that the values ​​of one column of the weighting matrix are not the same as the values ​​of other columns of the weighting matrix or a constant multiple of the values ​​of other columns of the weighting matrix.

[0036] In the weight matrix generation method described with reference to FIG. 5, the values ​​of each element of the weight matrix are set randomly, so the value of one column of the weight matrix may be the same as the value of another column or a constant multiple of the value of another column. In this case, deep learning performs a normalization process, so even if the weights are multiplied by a constant, they are treated as the same feature. Therefore, even if learning is performed using features based on the same weight pattern or a weight pattern obtained by multiplying a certain weight pattern by a constant, the effect of improving the prediction accuracy of the model is small.

[0037] On the other hand, in the weight matrix generation method described with reference to Figure 6, as described above, a weight matrix is ​​generated so that the values ​​in one column of the weight matrix are not the same as the values ​​in other columns or a constant multiple of the values ​​in other columns. This eliminates data processing that has little effect on improving the prediction accuracy of the model, enabling more efficient model learning.

[0038] 3 again, the statistical feature matrix generation unit 14 acquires a feature matrix including date and time of a preset period and feature data from the storage unit 12. The statistical feature matrix generation unit 14 generates a statistical feature matrix by multiplying the acquired feature matrix by the weight matrix generated by the weight matrix generation unit 13 (step S3). As shown in FIG. 9, the statistical feature matrix generation unit 14 generates an N row × L column (3 row × 5 column) statistical feature matrix by multiplying the N row × M column (3 row × 3 column) feature matrix shown in FIG. 2 by the M row × L column (3 row × 5 column) weight matrix shown in FIG. 4. The statistical feature matrix generation unit 14 stores the generated statistical feature matrix in the storage unit 12.

[0039] The feature creation device 10 performs the above-described processes from step S1 to step S3 at any timing. For example, the feature creation device 10 may perform the above-described processes when it acquires data (traffic data and feature data). Furthermore, the feature creation device 10 may perform the above-described processes, for example, before performing traffic prediction.

[0040] In the learning phase, the prediction function unit 15 acquires the statistical feature matrix and traffic data stored in the storage unit 12. The prediction function unit 15 learns a model for predicting the prediction target based on data of the prediction target (traffic in this embodiment) and features (second features) obtained from the statistical feature matrix (step S4). Specifically, the prediction function unit 15 sets the value of each column of the statistical feature matrix as a feature (second feature) and learns the model by deep learning. For deep learning, for example, the algorithm described in Non-Patent Document 1 may be used.

[0041] In the inference phase, feature data for any timing to be predicted is stored in the storage unit 12, and the prediction function unit 15 acquires the feature data. The weight matrix generation unit 13 and the statistical feature generation unit 14 perform statistical processing of the features (to generate statistical features) in the same manner as in steps S2 and S3. Here, the weight matrix used is the same as the weight matrix generated in the learning phase. The prediction function unit 15 predicts the prediction target (traffic) using the model learned in the learning phase (step S5).

[0042] As described above, the feature creation device 10 according to this embodiment includes an interface 11 as an acquisition unit, a weight matrix generation unit 13, and a statistical feature matrix generation unit 14 as a feature creation unit. The interface 11 acquires data of a prediction target and features (first features) related to the prediction target. The weight matrix generation unit 13 generates a weight matrix having the same number of rows as the number of columns of a feature matrix generated based on the acquired first features, in which the first features for each row element are arranged in the column direction, and the number of columns is greater than the maximum number of column elements expected for each row element. The statistical feature matrix generation unit 14 creates second features based on the statistical feature matrix, which is the product of the feature matrix and the weight matrix.

[0043] By generating a weight matrix corresponding to various weight patterns and creating a feature (second feature) based on a statistical feature matrix, which is the product of the weight matrix and the feature matrix, it is possible to create multiple feature values ​​with varying weights without searching for optimal weights (numerical transformation), as in the past. By creating multiple feature values ​​with varying weights, it is possible to combine the feature values ​​with varying weights and identify optimal weights through deep learning. Furthermore, according to the present disclosure, special skills of an engineer are not required to search for optimal weights. Therefore, according to the present disclosure, even when the correlation between the prediction target and the feature value is low, it is possible to more easily create feature values ​​through numerical transformation.

[0044] Next, the hardware configuration of the feature creation device 10 according to this embodiment will be described.

[0045] FIG. 10 is a diagram showing an example of the hardware configuration of a feature creation device 10 according to this embodiment. FIG. 10 shows an example of the hardware configuration of the feature creation device 10 when the feature creation device 10 is configured by a computer capable of executing program instructions. Here, the computer may be a general-purpose computer, a dedicated computer, a workstation, a PC (Personal Computer), an electronic notepad, or the like. The program instructions may be program code, code segments, or the like for performing necessary tasks.

[0046] 10, the feature creation device 10 includes a processor 21, a read-only memory (ROM) 22, a random access memory (RAM) 23, a storage 24, an input unit 25, a display unit 26, and a communication interface (I / F) 27. Each component is connected to each other via a bus 29 so as to be able to communicate with each other. The processor 21 is specifically a central processing unit (CPU), a micro processing unit (MPU), a graphics processing unit (GPU), a digital signal processor (DSP), a system on a chip (SoC), or the like, and may be configured by multiple processors of the same type or different types.

[0047] The processor 21 is a control unit that controls each component and performs various arithmetic operations. That is, the processor 21 reads a program from the ROM 22 or the storage 24 and executes the program using the RAM 23 as a work area. The processor 21 controls each component and performs various arithmetic operations in accordance with the program stored in the ROM 22 or the storage 24. In this embodiment, the ROM 22 or the storage 24 stores a program for operating a computer as the feature creation device 10 according to the present disclosure. The program is read and executed by the processor 21 to implement each component of the feature creation device 10, such as the weight matrix generation unit 13, the statistical feature matrix generation unit 14, and the prediction function unit 15.

[0048] The program may be provided in a form stored on a non-transitory storage medium such as a CD-ROM (Compact Disk Read Only Memory), a DVD-ROM (Digital Versatile Disk Read Only Memory), a USB (Universal Serial Bus) memory, etc. The program may also be provided in a form downloaded from an external device via a network.

[0049] The ROM 22 stores various programs and various data. The RAM 23 temporarily stores programs or data as a working area. The storage 24 is configured with an HDD (Hard Disk Drive) or an SSD (Solid State Drive) and stores various programs including an operating system and various data. The memory unit 12 is configured with, for example, any one of the ROM 22, the RAM 23, and the storage 24, or a combination of these.

[0050] The input unit 25 includes a pointing device such as a mouse and a keyboard, and is used to input various types of information.

[0051] The display unit 26 is, for example, a liquid crystal display, and displays various information. The display unit 26 may function as the input unit 25 by adopting a touch panel system.

[0052] The communication interface 27 is an interface for communicating with other devices (for example, the traffic collecting device 1 and the feature data source 2), and is, for example, an interface for a LAN. The interface 11 is configured by the communication interface 27, for example.

[0053] A computer can be suitably used to function as each unit of the feature creation device 10 described above. Such a computer can be realized by storing a program describing the processing to realize the functions of each unit of the feature creation device 10 in a storage unit of the computer, and having the processor of the computer read and execute the program. In other words, the program causes the computer to function as the feature creation device 10 described above. The program can also be stored in a non-transitory storage medium. The program can also be provided via a network.

[0054] The following additional notes are provided regarding the above-described embodiments.

[0055] [Additional note 1] A feature creation device that creates features related to a target to be predicted by a deep learning model, the feature creation device being used for learning the model using deep learning, Memory and a control unit connected to the memory; Equipped with The control unit acquiring data of the prediction target and a first feature amount related to the prediction target; generating a weight matrix having a number of rows equal to a number of columns of a feature matrix generated based on the acquired first feature, in which the first feature for one row element is arranged in a column direction, and having a number of columns greater than a maximum number of column elements assumed for the one row element; a feature creation device that creates a second feature based on a statistical feature matrix that is a product of the feature matrix and the weight matrix;

[0056] [Additional note 2] In the feature creation device according to Supplementary Item 1, The control unit generates the weighting matrix so that a value in one column of the weighting matrix is ​​not the same as a value in another column of the weighting matrix, and is not a constant multiple of a value in another column of the weighting matrix.

[0057] [Additional note 3] In the feature creation device according to Supplementary Item 1 or 2, The control unit learns the model based on the data to be predicted and the second feature.

[0058] [Additional note 4] 1. A feature creation method using a feature creation device that creates features related to a target to be predicted by a deep learning model, the feature creation method being used for learning the model using deep learning, the method comprising: acquiring data of the prediction target and a first feature amount related to the prediction target; generating a weight matrix having a number of rows equal to a number of columns of a feature matrix generated based on the acquired first feature, in which the first feature for one row element is arranged in a column direction, and having a number of columns greater than a maximum number of column elements assumed for the one row element; a feature creation method for creating a second feature based on a statistical feature matrix that is a product of the feature matrix and the weight matrix.

[0059] [Additional note 5] A non-transitory storage medium storing a program executable by a computer, the program causing the computer to operate as the feature creation device according to any one of claims 1 to 3.

[0060] Although the above-described embodiments have been described as typical examples, it will be apparent to those skilled in the art that many modifications and substitutions can be made within the spirit and scope of the present disclosure. Therefore, the present invention should not be interpreted as being limited by the above-described embodiments, and various modifications and alterations are possible without departing from the scope of the claims. For example, multiple building blocks shown in the block diagrams of the embodiments can be combined into one, or one building block can be divided. [Explanation of symbols]

[0061] 1 Traffic collection device 2. Feature Data Source 10 Feature creation device 11 Interface (acquisition part) 12 Storage section 13 Weight matrix generator 14 Statistical feature generation unit (feature creation unit) 15 Prediction Function Department 21 processors 22 ROM 23 RAM 24 Storage 25 Input section 26 Display section 27 Communication I / F 29 Bus

Claims

1. A feature creation device that creates features related to a target to be predicted by a deep learning model, the feature creation device being used for learning the model using deep learning, an acquisition unit that acquires data of the prediction target and a first feature amount related to the prediction target; a weight matrix generation unit that generates a weight matrix having a number of rows equal to a number of columns of a feature amount matrix in which the first feature amounts for one row element are arranged in a column direction, the feature amount matrix being generated based on the acquired first feature amounts, and the number of columns being greater than a maximum number of column elements assumed for the one row element; a feature creation unit that creates a second feature based on a statistical feature matrix that is the product of the feature matrix and the weight matrix.

2. 2. The feature creation device according to claim 1, the weight matrix generation unit generates the weight matrix so that a value in one column of the weight matrix is ​​not the same as a value in another column of the weight matrix and is not a constant multiple of a value in another column of the weight matrix.

3. 2. The feature creation device according to claim 1, a prediction function unit configured to train the model based on the data to be predicted and the second feature.

4. 1. A feature creation method using a feature creation device that creates features related to a target to be predicted by a deep learning model, the feature creation method being used for learning the model using deep learning, the method comprising: acquiring data of the prediction target and a first feature amount related to the prediction target; generating a weight matrix having the same number of rows as the number of columns of a feature matrix generated based on the acquired first feature, in which the first feature for one row element is arranged in a column direction, and the number of columns is greater than the maximum number of column elements assumed for the one row element; generating a second feature based on a statistical feature matrix that is a product of the feature matrix and the weight matrix.

5. A program that causes a computer to operate as the feature creation device according to claim 1.

Images