Feature quantity creation device, feature quantity creation method, and program
Patent Information
- Application Number
- PCT/JP2023/018141
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Filing Date
- 2023-05-15
- Publication Date
- 2024-11-21
Smart Images

Figure JP2023018141_21112024_PF_FP_ABST
Abstract
Description
Feature creation device, feature creation method, and program
[0001] The present disclosure relates to a feature creation device, a feature creation method, and a program.
[0002] When designing a communication network, it is possible to improve the efficiency of capital investment by calculating the required bandwidth for the physical links of the future communication network. By predicting (distribution estimation) the traffic generation probability distribution (traffic distribution), it is possible to estimate the required bandwidth with high accuracy (see Non-Patent Document 1).
[0003] Non-Patent Document 1 describes a technology that uses machine learning to estimate traffic distribution based on a feature vector consisting of traffic data and peak rate information for each flow, and calculates required bandwidth using the traffic distribution and future feature vector. In this technology, in the learning phase, a trained model is created based on feature vectors such as the past number of lines that have used the target link or the upper limit bandwidth set for each line. In the inference phase, traffic distribution is predicted using the feature vector for the situation (future) to be predicted and the created trained model.
[0004] E. Takeshita, T. Kosugi, T. Yoshida, “Traffic Statistical Upper Limit Prediction from Flow Features in Network Provisioning”, GLOBECOM 2021
[0005] In the technology described in Non-Patent Document 1, the accuracy of prediction is greatly affected by whether or not it is possible to create features that are highly correlated with the traffic distribution to be predicted. For example, in a link where the traffic volume per line is close to the upper limit bandwidth, there is a high correlation between the traffic and features created from the upper limit bandwidth. In this case, the technology described in Non-Patent Document 1 can achieve high prediction accuracy.
[0006] On the other hand, in networks where traffic varies greatly from line to line, such as networks where some lines carry heavy traffic, if the correlation between the features used for prediction and the traffic distribution is low, high prediction accuracy cannot be achieved. In other words, if a trained model is created without knowing what kind of traffic will flow on each line, prediction accuracy will decrease.
[0007] In consideration of the above-described problems, an object of the present disclosure is to provide a feature creation device, a feature creation method, and a program that can create features that are highly correlated with a prediction target and improve the prediction accuracy of the prediction target.
[0008] In order to solve the above problem, a feature creation device according to the present disclosure is a feature creation device that creates features related to a prediction target, and includes: a learning unit that creates a trained model that predicts the prediction target based on features related to the prediction target, using first data of the prediction target, whose features are known, and first features corresponding to the first data; a feature creation unit that creates second features related to the prediction target; a prediction unit that predicts the prediction target using the second features and the trained model; a comparison unit that compares the second data of the prediction target with a prediction result of the trained model and determines the accuracy of the prediction result of the trained model; and a feature adjustment unit that adjusts the value of the second feature until it is determined that the accuracy of the prediction result of the trained model is equal to or greater than a predetermined threshold.
[0009] 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 a prediction target, and includes the steps of: using first data of the prediction target, for which features are known, and first features corresponding to the first data, to create a trained model that predicts the prediction target based on the features related to the prediction target; creating second features related to the prediction target; predicting the prediction target using the second features and the trained model; comparing the second data of the prediction target with the prediction result of the trained model and determining the accuracy of the prediction result of the trained model; and adjusting the value of the second feature until it is determined that the accuracy of the prediction result of the trained model is equal to or greater than a predetermined threshold.
[0010] In order to solve the above problem, a program according to the present disclosure causes a computer to operate as the above-described feature creation device.
[0011] According to the feature creation device, feature creation method, and program disclosed herein, it is possible to create features that are highly correlated with a prediction target, thereby improving the prediction accuracy of the prediction target.
[0012] FIG. 4 is a diagram illustrating an example configuration of a feature creation device according to a first embodiment of the present disclosure. FIG. 5 is a flowchart illustrating an example of an operation of the feature creation device shown in FIG. 1. FIG. 6 is a flowchart illustrating an example of an operation when adjusting a feature in the feature creation device shown in FIG. 1. FIG. 7 is a diagram illustrating an example configuration of a feature creation device according to a second embodiment of the present disclosure. FIG. 8 is a flowchart illustrating an example of an operation when adjusting a feature in the feature creation device shown in FIG. 4. FIG. 9 is a diagram illustrating an example configuration of a computer that functions as the feature creation device according to the present disclosure.
[0013] Hereinafter, embodiments of the present disclosure will be described with reference to the drawings.
[0014] First Embodiment FIG. 1 is a diagram illustrating an example configuration of a feature creation device 100 according to a first embodiment of the present disclosure. The feature creation device 100 according to this embodiment creates features related to a prediction target and predicts the prediction target using the created features. The prediction target may be, for example, traffic distribution in a communication network, sales data for a product, or the number of customers. The following description uses an example in which the prediction target is traffic distribution in a communication network.
[0015] As shown in FIG. 1 , a feature creation device 100 according to this embodiment includes an interface 110, a recording unit 120, a learning unit 130, a feature creation unit 140, a prediction unit 150, a comparison unit 160, and a feature adjustment unit 170.
[0016] The interface 110 is connected to a training traffic data source 210, a training feature data source 220, an adjustment traffic data source 230, and a prediction feature data source 240, and transmits and receives traffic data and features thereto.
[0017] The training traffic data source 210 provides data on the traffic distribution to be predicted. The training traffic data source 210 provides, for example, traffic data, which is time-series data on the traffic volume at each date and time. The traffic data provided from the training traffic data source 210 is used for training a trained model, which will be described later. Hereinafter, the traffic data provided from the training traffic data source 210 will be referred to as training traffic data. The training traffic data as first data is data that combines distributions of known shapes. In other words, the training traffic data is data whose features are known.
[0018] The training feature data source 220 provides features corresponding to the training traffic data. The features provided from the training feature data source 220 are used for training a trained model, which will be described later. Hereinafter, the features provided from the training feature data source 220 will be referred to as training features. The training features as first features are features related to the distributions combined in the training traffic data, and indicate, for example, the types and number of combined distributions (e.g., which distributions and how many lines are combined at each date and time).
[0019] The adjustment traffic data source 230 provides data on the traffic distribution to be predicted. The adjustment traffic data source 230 provides, for example, traffic data, which is time-series data on the traffic volume at each date and time. The traffic data provided from the adjustment traffic data source 230 is used for adjusting feature quantities, which will be described later. Hereinafter, the traffic data provided from the adjustment traffic data source 230 will be referred to as adjustment traffic data. The adjustment traffic data as second data is, for example, past traffic data that has been actually measured.
[0020] The prediction feature data source 240 provides features used to predict traffic distribution, which is the target of prediction using a trained model. Hereinafter, the features provided from the prediction feature data source 240 will be referred to as prediction features. The prediction features as third features are created based on the features created by the feature creation device 100.
[0021] The interface 110 acquires training traffic data from the training traffic data source 210 and records it in the recording unit 120. The interface 110 also acquires training features from the training feature data source 220 and records it in the recording unit 120. The interface 110 also acquires adjustment traffic data from the adjustment traffic data source 230 and records it in the recording unit 120. The interface 110 also outputs features (adjusted second features) that have been adjusted by the feature adjustment unit 170 (described later) and that are recorded in the recording unit 120 to the prediction feature data source 240. The interface 110 also acquires prediction features from the prediction feature data source 240 and records them in the recording unit 120.
[0022] The recording unit 120 records the various features and prediction target data (traffic data) described above. The recording unit 120 also records a trained model created by the learning unit 130 (described later). The recording unit 120 also records prediction results based on the trained model.
[0023] The learning unit 130 reads out the training traffic data (first data) and training features (first features) recorded in the recording unit 120. The learning unit 130 uses the training traffic data and the training features to create a trained model that predicts a traffic distribution based on features related to the traffic distribution to be predicted. The learning unit 130 records the created trained model in the recording unit 120.
[0024] The feature creation unit 140 creates initial values of features (second features) related to the traffic distribution to be predicted. The values of the features created by the feature creation unit 140 are adjusted by the feature adjustment unit 170, which will be described later. Hereinafter, the feature to be adjusted is referred to as an adjustment feature. The feature creation unit 140 records the created adjustment feature in the recording unit 120.
[0025] The prediction unit 150 reads out the trained model and the feature values for adjustment recorded in the recording unit 120. The prediction unit 150 predicts a traffic distribution using the feature values for adjustment and the trained model. The prediction unit 150 records the prediction result of the trained model in the recording unit 120.
[0026] The comparison unit 160 reads the adjustment traffic data and the prediction result of the trained model (prediction result based on the adjustment features) recorded in the recording unit 120. The comparison unit 160 compares the adjustment traffic data (second data to be predicted) with the prediction result of the trained model, and determines the accuracy of the prediction result of the trained model. The comparison unit 160 outputs the determination result to the feature adjustment unit 170.
[0027] When the comparison unit 160 determines that the accuracy of the prediction result of the trained model is lower than a predetermined threshold, the feature adjustment unit 170 changes the value of the adjustment feature and records it in the recording unit 120. The feature adjustment unit 170 adjusts the value of the adjustment feature until it determines that the accuracy of the prediction result of the trained model is equal to or higher than the predetermined threshold.
[0028] Next, the operation of the feature creation device 100 according to this embodiment will be described.
[0029] FIG. 2 is a flowchart showing an example of the operation of the feature creation device 100 according to this embodiment, and is a diagram for explaining a feature creation method by the feature creation device 100.
[0030] The interface 110 acquires training traffic data from the training traffic data source 210 and training features from the training feature data source 220 (step S11). The interface 110 records the acquired training traffic data and training features in the recording unit 120. Training traffic data with known features can be created, for example, by using simulation. Training traffic data with known features can also be created, for example, by using traffic data of an application used in a communication network whose traffic distribution is to be predicted.
[0031] The learning unit 130 reads the training traffic data and training features recorded in the recording unit 120. The learning unit 130 uses the training traffic data (first data) and the training features (first features) corresponding to the training traffic data to create a trained model that predicts traffic distribution based on features related to the traffic distribution (step S12). Machine learning can be used to create the trained model. For example, the SVAE (Supervised-Variational AutoEncoder) described in Non-Patent Document 1 can be used as the machine learning.
[0032] The interface 110 acquires the adjustment traffic data from the adjustment traffic data source 230 (step S13), and records the acquired adjustment traffic data in the recording unit 120.
[0033] The feature creation unit 140 and the feature adjustment unit 170 create adjustment feature values and adjust their values (step S14). Adjustment of adjustment feature values will be described in detail with reference to FIG.
[0034] The feature creation unit 140 sets the feature identifier n to 1 (step S141).
[0035] The feature creation unit 140 sets an initial value Fnm as an adjustment feature for the traffic distribution Lm at the date and time tn (step S142). The feature creation unit 140 records the initial value Fnm of the adjustment feature in the recording unit 120.
[0036] The prediction unit 150 reads the adjustment features and the trained model recorded in the recording unit 120. The prediction unit 150 inputs the adjustment features into the trained model and obtains the traffic distribution pn at date and time tn (step S143). The prediction unit 150 records the prediction result of the trained model in the recording unit 120.
[0037] The comparison unit 160 reads the adjustment traffic data and the prediction result of the trained model recorded in the recording unit 120. The comparison unit 160 compares the adjustment traffic data with the prediction result of the trained model to determine the accuracy of the prediction result of the trained model. Specifically, the comparison unit 160 determines whether the difference between the traffic distribution pn and the traffic distribution at the date and time tn of the adjustment traffic data is equal to or less than a predetermined value (step S144). For example, the comparison unit 160 may evaluate whether the difference between the two traffic distributions is equal to or less than a predetermined value based on likelihood. Furthermore, for example, the comparison unit 160 may evaluate whether the difference between the two traffic distributions is equal to or less than a predetermined value based on the error of the parameters (e.g., mean and variance) of the two traffic distributions.
[0038] If it is determined that the difference between the two traffic distributions is not equal to or less than the predetermined value (step S144: No), the feature adjustment unit 170 changes the value of the adjustment feature (step S145) and records the result in the recording unit 120. The feature adjustment unit 170 adjusts the value of the adjustment feature so that the difference between the adjustment traffic data and the prediction result of the trained model is equal to or less than the predetermined value. That is, the feature adjustment unit 170 adjusts the value of the adjustment feature so that the difference between the traffic distribution of the adjustment traffic data at date and time tn and the traffic distribution pn predicted by the trained model is equal to or less than the predetermined value. The feature adjustment unit 170 may adjust the value of the adjustment feature using, for example, backpropagation. After the process of step S145, the prediction unit 150 performs the process of step S143.
[0039] If it is determined that the difference between the two traffic distributions is equal to or less than the predetermined value (step S144: Yes), the feature adjustment unit 170 determines the feature at the time when the difference between the two traffic distributions becomes equal to or less than the predetermined value as the adjusted feature for adjustment at date and time tn (step S146), and records it in the recording unit 120.
[0040] After the process of step S146, the feature creation unit 140 determines whether the identifier n is a specified maximum value (step S147).
[0041] If it is determined that the identifier n is equal to the specified maximum value (step S147: Yes), the adjustment of the adjustment feature is completed.
[0042] If it is determined that the identifier n is not the specified maximum value (step S147: No), the feature creation unit 140 increments the value of the identifier n by 1 (step S148) and returns to the processing of step S142.
[0043] In this embodiment, the adjustment of the feature value is repeated until the difference between the traffic distribution at date and time tn of the adjustment traffic data and the traffic distribution pn at date and time tn predicted by the learned model becomes less than a predetermined value.
[0044] 2 again, the interface 110 reads out the adjusted adjustment features recorded in the recording unit 120 and outputs them to the prediction feature data source 240. The prediction feature data source 240 creates prediction features based on the adjusted adjustment features. The prediction features are created, for example, by increasing or decreasing the values of the adjusted adjustment features by an arbitrary ratio.
[0045] The interface 110 acquires prediction features from the prediction feature data source 240 (step S15 ) and records them in the recording unit 120 .
[0046] The prediction unit 150 reads out the prediction features and the prediction model recorded in the recording unit 120. The prediction unit 150 inputs the prediction features into the trained model and predicts the traffic distribution (step S16).
[0047] As described above, the feature creation device 100 according to this embodiment includes a learning unit 130, a feature creation unit 140, a prediction unit 150, a comparison unit 160, and a feature adjustment unit 170. The learning unit 130 creates a trained model using training traffic data (first data to be predicted) whose features are known and training features (first features) corresponding to the training traffic data. The feature creation unit 140 creates adjustment features (second features) related to the prediction target. The prediction unit 150 predicts the prediction target using the adjustment features and the trained model. The comparison unit 160 compares the adjustment traffic data (second data to be predicted) with the prediction result of the trained model and determines the accuracy of the prediction result of the trained model. The feature adjustment unit 170 adjusts the value of the adjustment features until it is determined that the accuracy of the prediction result of the trained model is equal to or greater than a predetermined threshold.
[0048] This makes it possible to determine what combination of traffic distributions reproduces traffic and create features that correspond to actual traffic (features that are highly correlated with actual traffic).As a result, it is possible to avoid creating a trained model without knowing what kind of traffic will flow, thereby improving the prediction accuracy of the target.
[0049] Second Embodiment Fig. 4 is a diagram showing an example of the configuration of a feature creation device 100A according to a second embodiment of the present disclosure. In Fig. 4, the same components as those in Fig. 1 are denoted by the same reference numerals, and description thereof will be omitted.
[0050] 4 , the feature creation device 100A according to this embodiment includes an interface 110, a recording unit 120, a learning unit 130, a feature creation unit 140, a prediction unit 150, a comparison unit 160, a feature adjustment unit 170, and a residual determination unit 180. The feature creation device 100A according to this embodiment differs from the feature creation device 100 according to the first embodiment in that the residual determination unit 180 is added.
[0051] The residual determination unit 180 determines whether the number of feature adjustments made by the feature adjustment unit 170 has reached a predetermined number. If the residual determination unit 180 determines that the number of feature adjustments has reached the predetermined number, it notifies the outside via the interface 110.
[0052] 5 is a flowchart showing an example of the operation of adjusting features in the feature creation device 100A according to this embodiment. In FIG. 5, the same processes as those in FIG. 3 are denoted by the same reference numerals, and the description thereof will be omitted.
[0053] If it is determined that the difference between the traffic distribution pn at date and time tn predicted by the trained model and the traffic distribution at date and time tn of the adjustment traffic data is not less than a predetermined value (step S144: No), the residual determination unit 180 determines whether the number of times the feature adjustment unit 170 has adjusted the feature has reached a predetermined number (step S171).
[0054] If it is determined that the number of times the feature amount has been adjusted has not reached the predetermined number of times (step S171: No), the feature amount adjustment unit 170 proceeds to the process of step S145 and changes the value of the feature amount.
[0055] If it is determined that the number of feature adjustments has reached a predetermined number (step S171: Yes), the residual determining unit 180 notifies the device via the interface 110. For example, the residual determining unit 180 notifies that the difference between the traffic distribution pn at date and time tn predicted by the trained model and the traffic distribution at date and time tn of the adjustment traffic data is too large, and that the feature adjustment is not possible.
[0056] When the number of feature adjustments reaches a predetermined number, there is a possibility that the adjustment traffic cannot be reproduced by the training traffic and training features. In such a case, it is difficult to properly adjust the features, resulting in unnecessary calculations.
[0057] In this embodiment, by sending a notification when the number of feature adjustments reaches a predetermined number, it is possible to avoid unnecessary calculations by, for example, encouraging a review of the training traffic and training features.
[0058] The feature creation apparatuses 100 and 100A described above can each be realized by a computer 10 shown in FIG. 6 . A program for causing the computer 10 to function as the feature creation apparatuses 100 and 100A may be provided. The program may be stored on a storage medium or provided via a network. FIG. 6 is a block diagram showing a schematic configuration of a computer 10 functioning as the feature creation apparatuses 100 and 100A. The computer 10 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 executing necessary tasks.
[0059] 6, a computer 10 includes a processor 11, a ROM (Read Only Memory) 12, a RAM (Random Access Memory) 13, a storage 14, an input unit 15, a display unit 16, and a communication interface (I / F) 17. Each component is communicably connected to one another via a bus 19. The processor 11 is specifically a CPU (Central Processing Unit), an MPU (Micro Processing Unit), a GPU (Graphics Processing Unit), a DSP (Digital Signal Processor), a SoC (System on a Chip), or the like, and may be configured with multiple processors of the same type or different types.
[0060] The processor 11 is a control unit that controls each component and performs various arithmetic operations. That is, the processor 11 reads a program from the ROM 12 or the storage 14 and executes the program using the RAM 13 as a work area. The processor 11 controls each component and performs various arithmetic operations in accordance with the program stored in the ROM 12 or the storage 14. In this embodiment, the ROM 12 or the storage 14 stores a program for operating the computer 10 as the feature creation device 100, 100A according to the present disclosure. The processor 11 reads and executes the program to realize each component of the feature creation device 100, 100A described above.
[0061] 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.
[0062] The ROM 12 stores various programs and various data. The RAM 13 temporarily stores programs or data as a working area. The storage 14 is configured with an HDD (Hard Disk Drive) or an SSD (Solid State Drive) and stores various programs including the operating system and various data.
[0063] The input unit 15 includes a pointing device such as a mouse and a keyboard, and is used to input various types of information.
[0064] The display unit 16 is, for example, a liquid crystal display, and displays various information. The display unit 16 may be a touch panel type and function as the input unit 15.
[0065] The communication interface 17 is an interface for communicating with other devices, for example, an interface for a LAN.
[0066] The following additional notes are provided regarding the above-described embodiments.
[0067] [Supplementary Item 1] A feature creation device that creates features related to a prediction target, comprising a control unit, wherein the control unit: uses first data of the prediction target, for which features are known, and first features corresponding to the first data, to create a trained model that predicts the prediction target based on the features related to the prediction target; creates second features related to the prediction target; predicts the prediction target using the second features and the trained model; compares the second data of the prediction target with a prediction result of the trained model to determine the accuracy of the prediction result of the trained model; and adjusts the value of the second feature until it is determined that the accuracy of the prediction result of the trained model is equal to or greater than a predetermined threshold.
[0068] [Supplementary Item 2] In the feature creation device according to Supplementary Item 1, the control unit adjusts the value of the second feature so that a difference between the second data and a prediction result of the trained model is equal to or less than a predetermined value.
[0069] [Supplementary Item 3] In the feature creation device according to Supplementary Item 1 or 2, the control unit determines whether the number of times the value of the second feature has been adjusted has reached a predetermined number.
[0070] [Supplementary Item 4] In the feature creation device according to any one of Supplementary Items 1 to 3, the control unit predicts the prediction target using third features created based on the adjusted second features and the trained model.
[0071] [Supplementary Item 5] A feature creation method using a feature creation device that creates features related to a prediction target, the feature creation method comprising: using first data of the prediction target, whose features are known, and first features corresponding to the first data, to create a trained model that predicts the prediction target based on the features related to the prediction target; creating second features related to the prediction target; predicting the prediction target using the second features and the trained model; comparing the second data of the prediction target with the prediction result of the trained model to determine the accuracy of the prediction result of the trained model; and adjusting the value of the second feature until it is determined that the accuracy of the prediction result of the trained model is equal to or greater than a predetermined threshold.
[0072] [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 feature creation device according to any one of Supplementary Items 1 to 4.
[0073] 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.
[0074] REFERENCE SIGNS LIST 10 Computer 11 Processor 12 ROM 13 RAM 14 Storage 15 Input unit 16 Display unit 17 Communication I / F 19 Path 100, 100A Feature creation device 110 Interface 120 Recording unit 130 Learning unit 140 Feature creation unit 150 Prediction unit 160 Comparison unit 170 Feature adjustment unit 180 Residual determination unit
Claims
1. A feature creation device that creates features related to a prediction target, comprising: a learning unit that uses first data of the prediction target, whose features are known, and a first feature corresponding to the first data, to create a trained model that predicts the prediction target based on the features related to the prediction target; a feature creation unit that creates second features related to the prediction target; a prediction unit that predicts the prediction target using the second feature and the trained model; a comparison unit that compares the second data of the prediction target with a prediction result of the trained model and determines accuracy of the prediction result of the trained model; and a feature adjustment unit that adjusts the value of the second feature until it is determined that the accuracy of the prediction result of the trained model is equal to or greater than a predetermined threshold.
2. A feature creation device as described in claim 1, wherein the feature adjustment unit adjusts the value of the second feature so that the difference between the second data and the prediction result of the trained model is equal to or less than a predetermined value.
3. The feature creation device according to claim 1, further comprising a residual determination section which determines whether the number of times the value of the second feature has been adjusted by the feature adjustment section has reached a predetermined number.
4. A feature creation device as described in claim 1, wherein the prediction unit predicts the prediction target using a third feature created based on the adjusted second feature and the trained model.
5. A feature creation method by a feature creation device that creates features related to a prediction target, comprising the steps of: creating a trained model that predicts the prediction target based on features related to the prediction target, using first data of the prediction target, whose features are known, and first features corresponding to the first data; creating second features related to the prediction target; predicting the prediction target using the second features and the trained model; comparing the second data of the prediction target with the prediction result of the trained model and determining the accuracy of the prediction result of the trained model; and adjusting the value of the second feature until it is determined that the accuracy of the prediction result of the trained model is equal to or greater than a predetermined threshold.
6. A program for causing a computer to operate as the feature creation device according to claim 1.
Citation Information
Patent Citations
Learning device and learning method
JP2018088080A
Learning data set creation support device, and learning data set creation support method
JP2021179858A
Program, information processing device, and information processing method
JP2022016277A
Regularization of machine learning models
US20170161640A1
