Learning data generation device

Abstract

A learning data generation device (10) according to the present disclosure is provided with: an importance data generation unit (131) that, on the basis of related data, which is time-series data including an element having a correlation with a time-series change in target data, which is time-series data to be predicted, generates importance data indicating the importance of each arbitrary section of the target data; a sampling probability data generation unit (132) that, on the basis of the importance data, generates sampling probability data indicating the sampling probability of the target data for each arbitrary section; and a sampling unit (133) that, on the basis of the sampling probability data, samples the target data with the sampling probability for each arbitrary section to generate learning data.

Description

Learning data generation device

[0001] The present disclosure relates to a training data generation device.

[0002] Conventionally, training data used in machine learning to predict a target (e.g., future traffic in a communication network) has been generated by randomly sampling target data (e.g., past traffic data), which is time-series data of the target.

[0003] Furthermore, Non-Patent Document 1 describes a technique for determining the importance of each piece of data based on the amount of data change in any time series data, and determining the sampling probability of the time series data based on the determined importance.

[0004] Luis P. Silvestrin and Leonardos Pantiskas and Mark Hoogendoorn,"A Framework for Imbalanced Time-series Forecasting",2021

[0005] 8 is a diagram showing an example of time-series changes in target data. As shown in Fig. 8, learning data is generated from the target data of the prediction target during a learning period, and a prediction model constructed by machine learning of the generated learning data is used to predict the prediction target during a prediction period after the learning period.

[0006] Here, assume that during a learning period, settings (communication network settings (e.g., presence or absence of detouring traffic, total number of users, etc.)) that affect the prediction target (e.g., traffic in a communication network) are changed in the order of settings A, B, and C. Also assume that the period in which setting A is applied is greater than the period in which setting B is applied and greater than the period in which setting C is applied. In this case, an imbalance occurs in the number of observations (number of data items) in the target data during the period in which setting A is applied, the number of observations during the period in which setting B is applied, and the number of observations during the period in which setting C is applied. When such an imbalance in the number of observations occurs, a method of randomly sampling data may result in a higher number of samples during periods in which the number of observations is large (in the example shown in FIG. 8 , the period in which setting A is applied) and a lower number of samples during periods in which the number of observations is small (in the example shown in FIG. 8 , the period in which setting C is applied).

[0007] As a result, it becomes difficult to learn data from periods with few observations, and as shown in Figure 8, during the prediction period, the prediction accuracy may be high in periods when setting A' (setting A or a setting similar to setting A) is applied, while the prediction accuracy may be low in periods when setting C' (setting C or a setting similar to setting C) is applied.

[0008] Even with the technology described in Non-Patent Document 1, it is difficult to accurately determine the importance of each piece of data based solely on the amount of data change in the time-series data, and there is a problem in that prediction accuracy decreases when there is an imbalance in the number of observed data.

[0009] In view of the above circumstances, an object of the present disclosure is to provide a determination device that can generate learning data to be used in machine learning for predicting a prediction target, thereby enabling improvement in prediction accuracy.

[0010] In order to solve the above problem, a training data generation device according to the present disclosure is a training data generation device that generates training data to be used in machine learning for predicting a prediction target from target data, which is time-series data of the prediction target, and includes: an importance data generation unit that generates importance data indicating the importance of each arbitrary interval of the target data, based on associated data, which is time-series data including elements correlated with time-series changes in the target data; a sampling probability data generation unit that generates sampling probability data indicating the sampling probability of the target data for each arbitrary interval, based on the importance data; and a sampling unit that generates the training data by sampling the target data at the sampling probability for each arbitrary interval, based on the sampling probability data.

[0011] According to the training data generation device according to the present disclosure, it is possible to generate training data to be used in machine learning for predicting a prediction target, which can improve prediction accuracy.

[0012] 1 is a diagram illustrating an example of a configuration of a training data generation device according to an embodiment of the present disclosure. FIG. 2 is a flowchart illustrating an example of operation of the training data generation device illustrated in FIG. 1. FIG. 3 is a diagram illustrating an example of related data. FIG. 4 is a diagram illustrating another example of related data. FIG. 5 is a diagram illustrating an example of importance data generated by the importance data generation unit illustrated in FIG. 1 based on the related data shown in FIG. 4A. FIG. 6 is a diagram illustrating yet another example of related data. FIG. 7 is a diagram illustrating an example of importance data generated by the importance data generation unit illustrated in FIG. 1 based on the related data shown in FIG. 6A. FIG. 8 is a diagram illustrating an example of sampling probability data generated by the sampling probability data generation unit illustrated in FIG. 1. FIG. 9 is a diagram illustrating an example of target data.

[0013] 1 is a diagram illustrating an example configuration of a training data generation device 10 according to an embodiment of the present disclosure. The training data generation device 10 according to the present disclosure generates training data to be used in machine learning to predict a prediction target from target data, which is time-series data of the prediction target. The generated training data is output to a learning device 20 that constructs a prediction model that predicts the prediction target through machine learning of the training data. The prediction target may be, for example, communication network traffic, but is not limited to this, and various targets that change over time may be used as the prediction target.

[0014] 1, a training data generation device 10 according to this embodiment includes an input unit 11, a storage unit 12, a control unit 13, and an output unit 14. The input unit 11, the storage unit 12, the control unit 13, and the output unit 14 are connected via a data bus 15.

[0015] The input unit 11 includes one or more input interfaces that receive input operations from a user and acquire information based on the user operations. For example, the input unit 11 may be, but is not limited to, a pointing device, a keyboard, a mouse, or the like.

[0016] The input unit 11 accepts input of target data, which is time-series data to be predicted. The input unit 11 also accepts input of related data, which is time-series data including elements correlated with time-series changes in the target data. When the prediction target is traffic in a communication system, the related data is, for example, data for setting parameters of the communication network. The communication network parameters include, for example, the presence or absence of detouring traffic and the total number of users using the communication system, but are not limited to these, and may be any parameters that affect traffic in the communication network.

[0017] The input unit 11 outputs the input target data and related data to the storage unit 12 .

[0018] The storage unit 12 is one or more memories. The memory may be, for example, a semiconductor memory, a magnetic memory, or an optical memory, but is not limited to these, and may be any memory. The storage unit 12 is, for example, built into the training data generation device 10, but may also be configured to be accessible from outside the training data generation device 10 via any interface.

[0019] The storage unit 12 stores, for example, target data and related data output from the input unit 11. The storage unit 12 also stores importance data output from an importance data generation unit 131 (described later). The storage unit 12 also stores sampling probability data output from a sampling probability data generation unit 132 (described later). The storage unit 12 also stores learning data output from a sampling unit 133 (described later).

[0020] The control unit 13 is one or more processors. The processor may be, for example, a general-purpose processor or a dedicated processor specialized for a specific process, but is not limited to these and may be any processor. The control unit 13 controls the overall operation of the training data generation device 10.

[0021] The training data generation device 10 has the following software configuration: One or more programs used to control the operation of the training data generation device 10 are stored in a storage unit 12. When the programs stored in the storage unit 12 are read by a control unit 13, the programs cause the control unit 13 to function as an importance data generation unit 131, a sampling probability data generation unit 132, and a sampling unit 133.

[0022] The importance data generation unit 131 acquires the associated data stored in the storage unit 12. As described above, the associated data is time-series data including elements that have a correlation with time-series changes in the target data. The importance data generation unit 131 generates importance data indicating the importance of each arbitrary section of the target data based on the acquired associated data. The importance data generation unit 131 outputs the generated importance data to the storage unit 12. The generation of importance data by the importance data generation unit 131 will be described in detail later.

[0023] The sampling probability data generation unit 132 acquires importance data stored in the storage unit 12. Based on the acquired importance data, the sampling probability data generation unit 132 generates sampling probability data indicating the sampling probability of target data for each of the above-mentioned arbitrary sections. The sampling probability data generation unit 132 outputs the generated sampling probability data to the storage unit 12. Details of the generation of sampling probability data by the sampling probability data generation unit 132 will be described later.

[0024] The sampling unit 133 acquires sampling probability data and target data stored in the storage unit 12. Based on the acquired sampling probability data, the sampling unit 133 samples the target data at the sampling probability for each arbitrary section to generate training data. The sampling unit 133 outputs the generated training data to the storage unit 12.

[0025] The output unit 14 includes one or more output interfaces. The output unit 14 acquires the learning data stored in the storage unit 12 and outputs the learning data to the learning device 20.

[0026] Next, an operation of the training data generation device 10 according to this embodiment will be described. Fig. 2 is a flowchart showing an example of the operation of the training data generation device 10 according to this embodiment, and is a diagram for explaining a training data generation method executed by the training data generation device 10.

[0027] The input unit 11 acquires the input target data and related data (step S11) and stores them in the storage unit 12. The importance data generation unit 131 acquires the related data stored in the storage unit 12. The importance data generation unit 131 generates importance data indicating the importance of each arbitrary section of the target data based on the acquired related data (step S12). The importance data generation unit 131 stores the generated importance data in the storage unit 12.

[0028] The sampling probability data generation unit 132 acquires the importance data stored in the storage unit 12. Based on the acquired importance data, the sampling probability data generation unit 132 generates sampling probability data indicating the sampling probability of the target data for each of the arbitrary sections (step S13). The sampling probability data generation unit 132 stores the generated sampling probability data in the storage unit 12.

[0029] The sampling unit 133 acquires the sampling probability data and the target data stored in the storage unit 12. Based on the acquired sampling probability data, the sampling unit 133 samples the target data at the sampling probability for each arbitrary section to generate training data (step S14). The sampling unit 133 stores the generated training data in the storage unit 12.

[0030] Next, a description will be given of the generation of importance data by the importance data generator 131. In the following, it is assumed that the prediction target is traffic in a communication network, and the related data is data for setting parameters of the communication network (network setting data).

[0031] 3A is a diagram showing an example of network setting data. The network setting data includes parameters (elements) of a communication network for each predetermined time interval (every 5 minutes in the example shown in FIG. 3A). As shown in FIG. 3A, the network setting data may include multiple elements (elements 1, 2, ...).

[0032] The importance data generator 131 may generate importance data indicating the importance of a section in accordance with the length of the section in which the change in an element included in the network setting data (related data) is equal to or less than a predetermined value. In the example shown in Fig. 3A, the values ​​of elements 1 and 2 are constant in the section from 2023 / 01 / 01 00:00 to 2023 / 01 / 01 23:55.

[0033] FIG. 3B illustrates an example of importance data generated by the importance data generator 131 based on the network setting data illustrated in FIG. 3A . As described above, the values ​​of elements 1 and 2 are constant (the change in the elements is equal to or less than a predetermined value) in the interval from 2023 / 01 / 01 00:00 to 2023 / 01 / 01 23:55. In this case, the importance data generator 131 sets a constant value as the importance for the interval from 2023 / 01 / 01 00:00 to 2023 / 01 / 01 23:55, as illustrated in FIG. 3B . Here, the importance data generator 131 sets, for example, the reciprocal of the number of data points in each of the intervals in which the change in the element value is equal to or less than a predetermined value as the importance for each of the intervals. By doing so, as described with reference to FIG. 8 , even if there is an imbalance in the number of observations of target data in the intervals to which different settings are applied, the importance of the intervals to which each setting is applied can be equalized.

[0034] FIG. 4A is a diagram illustrating another example of network setting data. In the example shown in FIG. 4A, the value of element 1 at 23:55 on January 1, 2023, is significantly larger than the values ​​of element 1 at other times. In this way, when a value of a certain element significantly deviates from other values ​​of that element, the value is considered to be a value (outlier) resulting from a unique setting that occurs rarely. If importance data is generated taking such outliers into account, the importance of the interval corresponding to the outlier may be set high, potentially resulting in a decrease in the prediction accuracy of a prediction model constructed by machine learning of training data.

[0035] Therefore, as shown in FIG. 4B , if an element value (in the example shown in FIG. 4A , the value of element 1 at 23:55 on 2023 / 01 / 01) is an outlier, the importance data generation unit 131 lowers the importance of the interval corresponding to the element (the importance of the interval corresponding to 2023 / 01 / 01 23:55) compared to the importance of other intervals. This reduces the influence of a small number of idiosyncratic values. Note that the importance data generation unit 131 determines, for example, that a value whose difference from the average value of a certain element is equal to or greater than a predetermined value is an outlier.

[0036] FIG. 5A is a diagram showing yet another example of network setting data. The input unit 11 may accept a user's designation of an element from among multiple elements included in the network setting data (associated data). In the example shown in FIG. 5A, it is assumed that element 1 is designated. In this case, the importance data generation unit 131 generates importance data based on the value of the designated element 1 from among the multiple elements included in the network setting data (associated data), as shown in FIG. 5B. In this way, even when the associated data includes multiple elements, importance data can be generated using only the elements designated by the user (e.g., elements that the user considers to be particularly important).

[0037] 6A is a diagram showing yet another example of network setting data. As described above, the network setting data (related data) may include multiple elements (elements 1, 2, ...). In this case, the importance data generator 131 may set the importance of each interval (every 5 minutes in the example shown in FIGS. 6A and 6B) based on the values ​​of elements 1 and 2, as shown in FIG. 6B.

[0038] If the value of element 1 at time t is x_{1, t}, the value of element 2 at time t is x_{2, t}, and the importance at time t is w_t, the importance data generator 131 calculates the importance w_t at time t, for example, based on the following formula (1): w_t=f(x_{1, t+Δ}, x_{2, t+Δ})-f(x_{1, t}, x_{2, t}) Formula (1)

[0039] Here, the function f may be, for example, any linear function with the value x_{1, t} of element 1 and the value x_{2, t} of element 2 as variables. Also, the function f may be, for example, any nonlinear function such as a sigmoid function or a softplus function with the value x_{1, t} of element 1 and the value x_{2, t} of element 2 as variables.

[0040] Furthermore, the importance data generating unit 131 calculates the importance w_t at time t based on, for example, the following equation (2): w_t=p(x_{1, t}, x_{2, t}) Equation (2)

[0041] Here, the function p may be any function such as a softmax function or a normal distribution function that calculates the probability of occurrence of the value x_{1, t} of element 1 and the value x_{2, t} of element 2.

[0042] Note that the above-mentioned function is merely an example, and any function may be used to calculate the importance w_t as long as it can improve the prediction accuracy of the prediction model constructed by machine learning of the training data generated by the training data generation device 10.

[0043] Next, we will explain generation of sampling probability data by the sampling probability data generation unit 132. For example, if the importance data is data indicating the importance of five-minute intervals as shown in Figures 3B, 4B, 5B, and 6B, the sampling probability data generation unit 132 calculates the sampling probability for each interval, the same as the sampling probability data, as shown in Figure 7.

[0044] The sampling probability data generation unit 132, for example, normalizes the importance w_t to the range [0, 1], and calculates the sampling probability p_t at time t based on the following equation (3) using the normalized importance w_t at time t.

[0045]

[0046] Furthermore, the sampling probability data generating unit 132 applies an exponential function to the importance data, for example, and calculates a sampling probability p_t at time t based on the following equation (4).

[0047]

[0048] Here, α is an adjustment parameter, and the larger α is, the more likely it is that data with a high importance w_t will be selected.

[0049] As described above, the training data generation device 10 according to this embodiment includes an importance data generation unit 131, a sampling probability data generation unit 132, and a sampling unit 133. The importance data generation unit 131 generates importance data indicating the importance of each arbitrary section of target data, based on associated data, which is time-series data including elements correlated with time-series changes in target data, which is time-series data to be predicted. The sampling probability data generation unit 132 generates sampling probability data indicating the sampling probability of the target data for each arbitrary section, based on the importance data. The sampling unit 133 generates training data by sampling the target data at the sampling probability for each arbitrary section, based on the sampling probability data.

[0050] By generating importance data based on related data including elements correlated with time-series changes in the target data of the prediction target and generating sampling probability data based on the generated importance data, it is possible to generate importance data that is highly correlated with changes in the prediction target. Then, by generating training data by sampling the target data based on the sampling probabilities based on the importance data, it is possible to generate training data that can improve the prediction accuracy of the prediction target even when there is an imbalance in the number of data points in each element correlated with changes in the prediction target.

[0051] As described above, the program that controls the overall operation of the training data generation device 10 may be stored in a computer-readable storage medium. Using such a storage medium, the program can be installed on a computer. Here, the storage medium on which the program is stored may be a non-transitory storage medium. The non-transitory storage medium is not particularly limited, and may be, for example, a CD-ROM, a DVD-ROM, or a USB (Universal Serial Bus) memory. Furthermore, the program may be downloaded from an external device via a network.

[0052] The following additional notes are provided regarding the above-described embodiments.

[0053] [Supplementary Item 1] A training data generation device that generates training data to be used in machine learning for predicting a prediction target from target data, which is time-series data of the prediction target, comprising a control unit, wherein the control unit is configured to: generate importance data indicating the importance of each arbitrary interval of the target data, based on associated data, which is time-series data including elements correlated with time-series changes in the target data; generate sampling probability data indicating a sampling probability of the target data for each arbitrary interval, based on the importance data; and generate the training data by sampling the target data at the sampling probability for each arbitrary interval, based on the sampling probability data.

[0054] [Supplementary Item 2] In the training data generation device according to Supplementary Item 1, the control unit generates importance data indicating the importance of a section in accordance with the length of the section in which the change in the element is equal to or less than a predetermined value.

[0055] [Supplementary Item 3] In the training data generation device according to Supplementary Item 1 or 2, when the value of the element is an outlier, the control unit lowers the importance of the section corresponding to the element below the importance of other sections.

[0056] [Supplementary Item 4] In the training data generation device according to any one of Supplementary Items 1 to 3, the related data includes a plurality of elements, and the control unit generates the importance data based on a specified element from among the plurality of elements.

[0057] [Supplementary Item 5] A training data generation method executed by a training data generation device that generates training data to be used in machine learning to predict a prediction target from target data, which is time-series data of the prediction target, the training data generation method comprising: generating importance data indicating the importance of each arbitrary interval of the target data based on associated data, which is time-series data including elements correlated with time-series changes in the target data; generating sampling probability data indicating a sampling probability of the target data for each arbitrary interval based on the importance data; and generating the training data by sampling the target data at the sampling probability for each arbitrary interval based on the sampling probability data.

[0058] [Supplementary Item 6] A non-transitory storage medium storing a program executable by a computer, the non-transitory storage medium storing the program causing the computer to operate as the training data generation device according to Supplementary Item 1.

[0059] All publications, patent applications, and technologies mentioned in this specification are herein incorporated by reference to the same extent as if each individual publication, patent application, and technology was specifically and individually indicated to be incorporated by reference.

[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 construed as being limited by the above-described embodiments, and various modifications or alterations are possible without departing from the scope of the claims.

[0061] REFERENCE SIGNS LIST 10 Learning data generation device 11 Input unit 12 Storage unit 13 Control unit 14 Output unit 20 Learning device 131 Importance data generation unit 132 Sampling probability data generation unit 133 Sampling unit

Claims

1. A learning data generation device that generates learning data to be used in machine learning to predict a prediction target from target data, which is time-series data of the prediction target, comprising: an importance data generation unit that generates importance data indicating the importance of each arbitrary interval of the target data based on associated data, which is time-series data including elements that have a correlation with time-series changes in the target data; a sampling probability data generation unit that generates sampling probability data indicating the sampling probability of the target data for each arbitrary interval based on the importance data; and a sampling unit that generates the learning data by sampling the target data at the sampling probability for each arbitrary interval based on the sampling probability data.

2. A learning data generation device according to claim 1, wherein the importance data generation unit generates importance data indicating the importance of a section in accordance with the length of the section in which the change in the element is equal to or less than a predetermined value.

3. A training data generation device according to claim 1, wherein the importance data generation unit, when the value of the element is an outlier, lowers the importance of the section corresponding to the element compared to the importance of other sections.

4. A learning data generation device according to claim 1, wherein the related data includes a plurality of elements, and the importance data generation unit generates the importance data based on a specified element from among the plurality of elements.

Images