JP6550783B2 - Data output method, data output program and data output device - Google Patents

Description

  The present invention relates to a data output method, a data output program, and a data output device.

  In machine learning, a prediction model is generated using learning data, and prediction is performed using the generated prediction model. For this reason, machine learning changes its performance according to learning data. Therefore, there is a technique for selecting data effective for learning. For example, the correlation between the input time-series data of a plurality of types of feature amounts and the time-series data of the objective variable is analyzed, and the feature amount having a high influence on the objective variable is specified.

Unexamined-Japanese-Patent No. 2012-27880

  However, conventional techniques may not be able to identify data effective for learning. Among events to be predicted, the occurrence principle changes depending on the time zone. On the other hand, in the prior art, feature quantities that are correlated with the target variable in the entire data are identified regardless of the generation principle. For this reason, in the related art, a feature amount having a high correlation with a target variable is not selected in a specific time zone, and there are cases in which data that is effective for learning can not be identified.

  In one aspect, it is an object of the present invention to provide a data output method, a data output program, and a data output device that can identify data effective for learning.

  In the first proposal, the data output method includes time-series data of values indicating the state of a specific monitoring target and time-series data of values indicating a plurality of monitoring target states different from the specific monitoring target. Calculate the correlation value of The data output method classifies time-series data indicating a plurality of monitoring target states into a plurality of clusters based on the correlation between the calculated correlation values. The data output method extracts any time series data from each of a plurality of clusters. The data output method outputs information indicating the type of each time-series data extracted from each of the plurality of clusters.

  According to one embodiment of the present invention, there is an effect that data effective for learning can be specified.

FIG. 1 is a diagram illustrating an example of a functional configuration of the data output apparatus according to the first embodiment. FIG. 2 is a diagram illustrating an example of the data configuration of the objective variable data. FIG. 3 is a diagram illustrating an example of a data configuration of feature amount data. FIG. 4 is a diagram for explaining an example of calculation of correlation values. FIG. 5A is a diagram for explaining an example of a window. FIG. 5B is a view for explaining an example of a window. FIG. 5C is a diagram for explaining an example of a window. FIG. 6 is a diagram illustrating an example of how to obtain a score. FIG. 7 is a diagram for explaining the feature amount of each cluster. FIG. 8 is a diagram showing an example of an event whose occurrence principle changes. FIG. 9 is a diagram showing changes in traffic volume on a certain road, and changes in time zone, number of accidents, and precipitation. FIG. 10 is a diagram illustrating an example of generating a prediction model for each time period. FIG. 11 is a diagram illustrating an example of generating a detailed prediction model in accordance with an event to be predicted. FIG. 12 is a diagram illustrating an example of objective variable data and feature amount data. FIG. 13 is a diagram illustrating an example of division of objective variable data and a plurality of feature amount data. FIG. 14 is a diagram showing an example of correlation calculation. FIG. 15 is a diagram illustrating an example of classifying data of a plurality of feature amounts. FIG. 16 is a diagram illustrating an example of a score for each cluster. FIG. 17 is a diagram illustrating an example of feature amount extraction for each cluster. FIG. 18 is a flowchart illustrating an example of the procedure of the data output process according to the first embodiment. FIG. 19 is a diagram illustrating an example of a functional configuration of the data output apparatus according to the second embodiment. FIG. 20 is a diagram illustrating an example of the correlation value of each feature amount with respect to the objective variable. FIG. 21 is a diagram illustrating an example in which the feature amount having the highest score is extracted. FIG. 22 is a diagram illustrating an example in which similar feature amounts are excluded from extraction targets. FIG. 23 is a diagram illustrating an example in which the feature amount having the highest score is extracted from the remaining feature amounts. FIG. 24 is a diagram illustrating an example in which similar feature amounts are excluded from extraction targets. FIG. 25 is a diagram illustrating an example in which the feature amount having the highest score is extracted from the remaining feature amounts. FIG. 26 is a flowchart of an example of a data output process according to the second embodiment. FIG. 27 is a diagram illustrating a computer that executes a data output program.

  Hereinafter, embodiments of a data output method, a data output program and a data output device according to the present invention will be described in detail based on the drawings. The present invention is not limited by this embodiment. And each Example can be suitably combined in the range which does not make processing contents contradictory.

[Device configuration]
A data output device 10 according to the present embodiment will be described. The data output device 10 is a device that identifies and outputs data effective for learning from various data that can be used for generating a prediction model of machine learning. The data output device 10 is, for example, a computer such as a personal computer or a server computer. The data output device 10 performs learning using data effective for learning to generate a prediction model, and performs prediction using the generated prediction model.

  FIG. 1 is a diagram illustrating an example of a functional configuration of the data output apparatus according to the first embodiment. As illustrated in FIG. 1, the data output device 10 includes a communication I / F (interface) unit 20, an input unit 21, a display unit 22, a storage unit 23, and a control unit 24. The data output device 10 may have other devices other than the above-described devices.

  The communication I / F unit 20 is an interface that controls communication with other devices. As the communication I / F unit 20, a network interface card such as a LAN card can be adopted.

  The communication I / F unit 20 transmits and receives various types of information to and from other devices via a network (not shown). For example, the communication I / F unit 20 receives various data used for generating a prediction model in machine learning. For example, the communication I / F unit 20 receives time-series data of values indicating the state of a specific monitoring target to be predicted by machine learning. The time-series data of values indicating the state of a specific monitoring target that is the target of prediction is data of an objective variable when generating a prediction model by machine learning. Further, the communication I / F unit 20 receives time-series data of values indicating the states of a plurality of monitoring targets different from the specific monitoring target. The time series data of values indicating the states of a plurality of monitoring targets different from the specific monitoring target are candidates for learning data when generating a prediction model by machine learning.

  The input unit 21 is an input device for inputting various types of information. The input unit 21 may be an input device that receives an input of an operation such as a mouse or a keyboard. The input unit 21 receives input of various types of information. For example, the input unit 21 receives various operation inputs related to machine learning. The input unit 21 receives an operation input from the user, and inputs operation information indicating the received operation content to the control unit 24.

  The display unit 22 is a display device that displays various information. Examples of the display unit 22 include display devices such as a liquid crystal display (LCD) and a cathode ray tube (CRT). The display unit 22 displays various information. For example, the display unit 22 displays various screens such as various operation screens and a screen showing a prediction result.

  The storage unit 23 is a storage device that stores various data. For example, the storage unit 23 is a storage device such as a hard disk, a solid state drive (SSD), or an optical disk. The storage unit 23 may be a semiconductor memory that can rewrite data, such as a random access memory (RAM), a flash memory, and a non-volatile static random access memory (NVSRAM).

  The storage unit 23 stores an OS (Operating System) executed by the control unit 24 and various programs. For example, the storage unit 23 stores various programs including programs for executing various processes described later. Furthermore, the storage unit 23 stores various data used in a program executed by the control unit 24. For example, the storage unit 23 stores objective variable data 30 and feature amount data 31.

  The target variable data 30 is data in which data of a target variable of machine learning is stored. In the objective variable data 30, time-series data of values indicating the state of a specific monitoring target to be predicted by machine learning is stored as data of the objective variable.

FIG. 2 is a diagram illustrating an example of the data configuration of the objective variable data. In the target variable data 30, a value indicating the state of a specific monitoring target is stored as data of the target variable for each measured time. In the example of FIG. 2, object variables x ₁ corresponds to the time t _1, the dependent variable x ₂ corresponds to the time t _2, · · ·, the target variable x _m corresponding to the time t _m stored .

  Returning to FIG. 1, the feature amount data 31 is data that stores data of a plurality of feature amounts that are candidates for learning data when a prediction model is generated by machine learning. The feature amount data 31 stores time-series data of values indicating the states of a plurality of monitoring targets different from a specific monitoring target to be predicted by machine learning as data of a plurality of feature amounts.

FIG. 3 is a diagram illustrating an example of a data configuration of feature amount data. In the feature amount data 31, values indicating the states of a plurality of feature amounts are stored as feature amount data for each measured time. In the example of FIG. 3, fa to fz indicating the type of the feature amount are stored as the type. Further, in the example of FIG. 3, the feature quantity Fa～fz, feature amount fa ₁ corresponds to the time t _1, fb ₁ ~fz _1, feature amounts fa ₂ corresponds to the time t _2, fb ₂ ~fz ₂ The feature quantities fa _m and fb _{m to} fz _m are stored corresponding to the time t _m .

  Returning to FIG. 1, the control unit 24 is a device that controls the data output device 10. As the control unit 24, an electronic circuit such as a CPU (Central Processing Unit) and an MPU (Micro Processing Unit), or an integrated circuit such as an ASIC (Application Specific Integrated Circuit) and an FPGA (Field Programmable Gate Array) can be employed. The control unit 24 has an internal memory for storing programs defining various processing procedures and control data, and executes various processes using these. The control unit 24 functions as various processing units by operating various programs. For example, the control unit 24 includes a reception unit 40, a calculation unit 41, a classification unit 42, an extraction unit 43, an output unit 44, and a prediction unit 45.

  The reception unit 40 performs various receptions. For example, the receiving unit 40 receives various operation instructions. For example, the reception unit 40 displays an operation screen related to machine learning on the display unit 22 and receives an operation instruction such as a process start from the input unit 21.

  The calculator 41 performs various calculations. For example, the calculation unit 41 calculates a correlation value between time-series data indicating a state of a specific monitoring target and each of time-series data indicating values of a plurality of monitoring targets different from the specific monitoring target. calculate. For example, the calculation unit 41 divides time-series data of values indicating a specific monitoring target state and time-series data of values indicating a plurality of monitoring target states into a plurality of periods. For example, the calculation unit 41 divides the objective variable data stored in the objective variable data 30 and the plurality of feature amount data stored in the feature amount data 31 into windows for each predetermined time period. Then, the calculation unit 41 calculates, for each period, a correlation value between time-series data of values indicating a specific monitoring target state and time-series data of values indicating a plurality of monitoring target states. For example, the calculation unit 41 calculates a correlation value between the objective variable and a plurality of feature amounts for each window. The correlation value may be calculated by any method as long as the value indicates the degree of correlation between the objective variable and the feature amount. For example, the calculation unit 41 calculates a product moment correlation coefficient as the correlation value.

FIG. 4 is a diagram for explaining an example of the calculation of the correlation value. In the example of FIG. 4, the window is set for each day of the week. Calculation unit 41, for each and purpose variables x ₁ ~x _m stored in the objective variable data 30, the feature quantity fa ₁ ~fa _m stored in the feature data 31, ..., and fz ₁ ~fz _m day Split into windows. Then, the calculation unit 41 calculates, for each window, a correlation value between the target variable x of the corresponding time and each of the plurality of feature quantities fa to fz. In the example of FIG. 4, the correlation value with the target variable x about the feature-value fa-fz is shown for every day of the week.

  In the example of FIG. 4, the case where the predetermined time zone is a day of the week and the window of each day is divided is illustrated. However, it is not limited to this. For example, the window duration may be specified by the user. For example, when the timing at which the occurrence principle of the event to be predicted changes is known, the user may designate the timing at which the occurrence principle changes as the window period. For example, it is assumed that the principle of occurrence of the event to be predicted changes depending on the time zone, day of the week, and month. In this case, the user designates a time zone, a day of the week, and a month according to the timing when the generation principle changes as the window period. The calculation unit 41 may divide the data of the target variable and the data of the feature amount for each designated time zone, day of the week, or month. FIG. 5A is a diagram for explaining an example of a window. In the example of FIG. 5A, the user designates the window period for each day of the week, and the calculation unit 41 divides the objective variable x and the feature amounts fa to fz into windows for each day of the week.

  Also, for example, the window period may be determined based on the change point of data. For example, a change point may occur in the data of the objective variable x and the feature amounts fa to fz depending on the generation principle of the event to be predicted. In this case, the calculation unit 41 may divide the data of the objective variable and the data of the feature amount for each change point of the data. FIG. 5B is a view for explaining an example of a window. In the example of FIG. 5B, change points of the data of the objective variable x and the feature amounts fa to fz are generated, and the calculation unit 41 displays the data of the objective variable x and the feature amounts fa to fz in a window for each change point. To divide. As long as the change point corresponds to the change in the generation principle, it may be any point such as a point that passes a threshold, a maximum point, or a minimum point. In FIG. 5B, a point that passes the threshold is a change point. By determining the window period based on the data change point in this way, the calculation unit 41 can divide the data in accordance with the change in the generation principle.

  Further, for example, the window periods may overlap each other. For example, the calculation unit 41 may divide the data for each window period while shifting the window by a period shorter than the certain period with the certain period as the window period. FIG. 5C is a diagram for explaining an example of a window. In the example of FIG. 5C, the calculation unit 41 divides the data of the objective variable x and the feature quantities fa to fz into windows while overlapping the window periods. By overlapping the window periods in this way, the calculation unit 41 can divide the feature amount data even when the change in the generation principle is not clear, and the similar feature amount can be the same by the classification unit 42 described later. It can be classified into classification.

Returning to FIG. 1, the classification unit 42 performs various classifications. For example, the classification unit 42 classifies time series data indicating the states of a plurality of monitoring targets into a plurality of clusters based on the correlation between correlation values. For example, the classification unit 42 classifies time-series data of values indicating the states of a plurality of monitoring targets into a plurality of clusters based on the distribution of correlation values calculated for each period in a plurality of periods. For example, the classification unit 42 classifies the feature quantities having similar changes in correlation for each window into the same class, and classifies the feature quantities into a plurality of clusters. For example, for each feature quantity, the classification unit 42 obtains an error in correlation values with other feature quantities in each period. For example, the classification unit 42 obtains the Euclidean distance of the correlation value of each period as the error of the correlation value in each period. When the correlation value of the feature quantity fa in the periods _{1 to m} is ta _{1 to} tam and the correlation value of the feature quantity fb is tb ₁ to tb _m , the Euclidean distance is the period as shown in the following equation (1). The difference between each correlation value ta and tb is squared and summed, and the square root of the total value is calculated.

Euclidean distance = ((ta ₁ −tb ₁ ) ² +... + (Ta _m −tb _m ) ² ) ^1/2 (1)

  The classification unit 42 classifies feature quantities close in Euclidean distance into the same classification. For example, the classifying unit 42 classifies feature quantities into clusters by repeatedly classifying feature quantities having a Euclidean distance equal to or less than a threshold value to the same classification with any feature quantity as a reference. In the example of FIG. 4, the feature amount fa and the feature amount fb are classified into cluster A, and the feature amount fc and the feature amount fd are classified into cluster B.

  The extraction unit 43 performs various types of extraction. For example, the extraction unit 43 extracts any time series data from each of the plurality of clusters. For example, for each cluster, the extraction unit 43 obtains a score by weighting the correlation value for each period calculated for the time-series data classified into the cluster with a predetermined weight. For example, the extraction unit 43 weights the correlation value of each feature amount for each window with a predetermined weight. Then, the extraction unit 43 obtains a score from the weighted correlation value for each cluster. For example, the extraction unit 43 obtains, as a score, an average value, a maximum value, or a minimum value of each correlation value of the weighted feature amount for each cluster. Then, the extraction unit 43 extracts, for each cluster, the feature amount having the largest score. In the example of FIG. 4, the feature quantity fa is extracted from the cluster A, and the feature quantity fd is extracted from the cluster B.

  Here, how to obtain the score will be described. FIG. 6 is a diagram illustrating an example of how to obtain a score. In FIG. 6, to simplify the explanation, an example is described in which three feature quantities fa, fb, and fc are classified into one cluster, and correlation values are calculated for the three windows W1, W2, and W3. Do. FIG. 6A shows correlation values in the windows W1, W2, and W3 for the feature amounts fa, fb, and fc. FIG. 6B shows the result of weighting the correlation value with three patterns A to C. Pattern A shows a case where the correlation values are weighted equally. For example, in the pattern A, “1” is weighted equally to the correlation value. In this case, the weighted correlation value is the same as in FIG. Pattern B is shown as being weighted more heavily as the correlation value is larger. For example, in the pattern B, a value obtained by squaring the correlation value is used as the correlation value after weighting. Pattern C shows the case where the correlation value is equal to or greater than the predetermined threshold value, and the correlation value is used as the weighted correlation value as it is, and when the correlation value is less than the predetermined threshold value, it is set to zero. In the pattern C shown in FIG. 6 (B), the threshold value is 0.5. In this case, the weighted correlation value is the same as FIG. 6A when the correlation value is 0.5 or more, and is zero when the correlation value is less than 0.5. Note that the weighting pattern is not limited to this.

  The extraction unit 43 obtains a score from the weighted correlation value for each cluster. FIG. 6B shows the result of obtaining the average value, maximum value, or minimum value of each correlation value for each of the patterns A to C as the score. The extraction unit 43 extracts a feature amount for each cluster based on the score. For example, the extraction unit 43 extracts, for each cluster, the feature amount having the largest score. For example, when extracting the feature quantity having the maximum average score value, the feature quantity fc is extracted from the pattern A, the feature quantity fb is extracted from the pattern B, and the feature quantity fb is extracted from the pattern C. When extracting the feature amount having the maximum score value, the feature amount fb is extracted from the pattern A, the feature amount fb is extracted from the pattern B, and the feature amount fb is extracted from the pattern C. When extracting the feature quantity having the largest minimum score value, the feature quantity fc is extracted from the pattern A, the feature quantity fc is extracted from the pattern B, and the feature quantity fc is extracted from the pattern C.

  When extracting the feature quantity having the maximum score average value, the feature quantity whose weighted correlation value is high on average in all windows is extracted. When extracting the feature quantity that maximizes the maximum score value, the feature quantity having the highest weighted correlation value in any window is extracted. That is, feature quantities that have a large influence on the target variable in a specific window are extracted. When extracting the feature quantity having the maximum minimum score value, the feature quantity without a window having a low weighted correlation value is extracted. That is, a feature amount that does not have a window having a small influence on the objective variable is extracted.

FIG. 7 is a diagram for explaining the feature amount of each cluster. In the example of FIG. 7, in order to simplify the description, the case of classifying into clusters using the correlation of two windows W1 and W2 will be described. In the example of FIG. 7, the vertical axis indicates the correlation between the objective variable of the window W2 and the feature quantity, and the horizontal axis indicates the correlation between the target variable of the window W1 and the feature quantity. In the example of FIG. 7, the respective feature quantities are plotted according to the correlation in the window W1 and the correlation in the window W2. The feature amounts are classified into clusters for each of those having a short Euclidean distance. The Euclidean distance is the distance between the feature points. In the example of FIG. 7, the feature amounts are classified into _four clusters C _{1 to} C ₄ . The cluster C ₁ is a classification of feature amounts having high correlation in the window W2 and low correlation in the window W1. The cluster C ₂ is a classification of feature amounts having high correlation between the window W1 and the window W2. The cluster C ₃ is a classification of feature amounts having a low correlation between the window W1 and the window W2. Cluster C ₄ has a high correlation window W1, the correlation window W2 is lower characteristic of classification.

The extraction unit 43 extracts a feature amount for each cluster based on the score. For example, the extraction unit 43, a feature amount f1 is extracted from the cluster C _1, the feature amount f2 extracted from the cluster C _2, the feature quantity f3 extracted from the cluster C _3, extracts a feature quantity f4 from the cluster C ₄ .

The output unit 44 performs various outputs. For example, the output unit 44 outputs information indicating the type of each time-series data extracted from each of the plurality of clusters. For example, the output unit 44 outputs information indicating the type of feature amount extracted for each cluster by the extraction unit 43. For example, in the case of FIG. 7, the output unit 44 outputs the feature amounts f1, f2, f3, and f4 as information indicating the type of the extracted feature amount. In the clusters C _{1 to} C ₄ , feature quantities having close correlation with the objective variable are classified. In machine learning, even if many feature quantities that have a similar correlation with the objective variable are used, only similar patterns are learned, so prediction accuracy is difficult to improve, and there is diversity in the correlation with the objective variable. Learning with is preferable. Therefore, by extracting and outputting the feature amount for each cluster, it is possible to extract data having various correlations with the objective variable and to identify data effective for learning, thereby improving the prediction accuracy of machine learning. be able to.

  The prediction unit 45 performs various predictions by machine learning. For example, the prediction unit 45 generates a prediction model using the time-series data of the type output by the output unit 44 as learning data. And the prediction part 45 performs prediction with the produced | generated prediction model.

  Here, among the events that machine learning targets for prediction, the generation principle may change depending on the time zone. FIG. 8 is a diagram showing an example of an event whose occurrence principle changes. FIG. 8 shows changes in traffic volume on weekdays and holidays on a certain road, and changes in precipitation. On weekday roads, there are many commuting vehicles. Commuters travel on the road regardless of precipitation. For this reason, the traffic volume on weekdays is unlikely to be affected by precipitation even during periods of heavy precipitation, as indicated by reference numeral 60 in FIG. On the other hand, there are many tourist (excursion) vehicles on holiday roads. Tourism is better if the weather is better. For this reason, there are many vehicles for sightseeing when the weather is good, and the more precipitation there is, the more it decreases. For this reason, as for the traffic volume of a holiday, as shown to the code | symbol 61 of FIG. 8, the traffic volume reduces in the period when there is much precipitation. Thus, the traffic volume of the road changes the generation principle of traffic volume on weekdays and holidays.

  For example, a case is assumed in which a feature quantity having a correlation with an objective variable is specified in the entire data regardless of the generation principle. In this case, feature quantities having high correlation with the target variable in a specific time zone can not be selected. FIG. 9 is a diagram showing changes in traffic volume on a certain road, and changes in time zone, number of accidents, and precipitation. In the example of FIG. 9, the traffic volume, the time zone, and the number of accidents are correlated with the entire data. For this reason, the correlation of the entire data is at a medium level. On the other hand, traffic volume and precipitation change depending on whether it is a weekday or a holiday, and the correlation is low as the whole data. For this reason, the correlation in the whole data is low level. In this case, in the prior art, the time zone and the number of incidents are specified as feature quantities that are correlated with the target variable. That is, in the conventional technique, only a feature quantity having a certain correlation is selected regardless of the data generation principle, and therefore, an effective feature quantity is overlooked in a specific event such as precipitation.

  On the other hand, the data output apparatus 10 according to the present embodiment classifies the feature amounts having similar correlation changes into clusters, and extracts the feature amounts for each cluster, so that it is effective for a specific event such as precipitation. Feature amounts can also be extracted. As described above, since the data output device 10 can specify data effective for learning, the prediction accuracy of machine learning can be improved.

  In addition, for example, when the occurrence principle of an event that machine learning predicts changes depending on the time zone, learning is performed for each time zone using feature amount data, and a prediction model is set for each time zone. Assume the case of generation. FIG. 10 is a diagram illustrating an example of generating a prediction model for each time period. FIG. 10 shows a change in traffic volume on a weekday and a holiday on a road. In the example of FIG. 10, weekday features are learned using weekday feature data to generate a weekday prediction model. In the example of FIG. 10, a holiday prediction model is generated by learning holiday features using holiday feature data. In this case, in order to correspond to the event to be predicted, a prediction model is finely generated in accordance with the event to be predicted. FIG. 11 is a diagram illustrating an example of generating a detailed prediction model in accordance with an event to be predicted. In the example of FIG. 11, learning is performed using data of feature quantities in a daytime daytime zone on a weekday, and a prediction model of daytime traffic volume on a weekday is generated. When the prediction model is generated in detail in accordance with the event to be predicted, feature amount data that can be used for the prediction model decreases. FIG. 11 shows the range of feature quantity data that can be used for a daytime traffic volume forecast model on weekdays. Thus, when the amount of feature data that can be used in the prediction model decreases, the prediction accuracy of the prediction model decreases.

  On the other hand, the data output apparatus 10 according to the present embodiment can predict with one prediction model even if the occurrence principle changes depending on the time zone, for which the event for which machine learning is the target of prediction changes. Further, the data output device 10 generates a prediction model by using all time series data of a type effective for learning as learning data. As a result, since the data output device 10 can secure data of feature amounts that can be used for the prediction model, it is difficult for a drop in prediction accuracy of the prediction model to occur due to a lack of data.

  Next, description will be made using a specific example. Below, the case where a prediction model of traffic is generated is explained to an example. FIG. 12 is a diagram illustrating an example of data of an objective variable and data of a feature amount. FIG. 12 shows traffic volume data for each measured time as data of the target variable. Further, FIG. 12 shows, as data of the feature amount, data of precipitation, temperature, communication amount, and electric energy for each measured time. Precipitation is the amount of precipitation that falls in the area where traffic was measured. The temperature is the temperature of the area where the traffic volume is measured. The communication amount is the communication amount in which communication is performed in the regional network including the area where the traffic amount is measured. The amount of electric power is the amount of electric power used in the area including the area where the traffic volume is measured. The feature amount data also includes secondary data generated from the measured data. FIG. 12 shows a moving average value of the air temperature and the communication amount two unit time ago as data of the feature amount. The temperature two units before is the temperature measured two times before. The moving average value of the communication amount is an average value of the communication amounts up to a predetermined time before.

  The calculation unit 41 divides the data of the objective variable and the data of the plurality of feature amounts into windows for each predetermined time zone. FIG. 13 is a diagram illustrating an example of division of objective variable data and a plurality of feature amount data. In the example of FIG. 13, the window is set for each day of the week. The calculation unit 41 divides data of the objective variable and data of a plurality of feature amounts into windows W1 to W3 for each day of the week.

  The calculation unit 41 calculates, for each window, correlation values between the data of the objective variable and the data of the plurality of feature amounts. For example, the calculation unit 41 calculates a product moment correlation coefficient as the correlation value. FIG. 14 is a diagram showing an example of correlation calculation.

Here, an example of calculation of a product moment correlation coefficient will be described. For the data X = (x ₁ ,..., X _n ), the mean, variance, and standard deviation of X can be expressed as the following formulas (2) to (4).

Average of X: μ (X) = (x ₁ +... + X _n ) / n (2)
Dispersion of X: σ ² (X) = {(x ₁ −μ (X)) ² +...
+ (X _n- μ (X)) ² } / n (3)
Standard deviation of X: σ (X) = (σ ² (X)) ^1/2 (4)

Also, for data X = (x ₁ ,..., X _n ) and data Y = (y ₁ ,..., Y _n ), the covariance of X and Y is given by the following equation (5) It can be expressed as

Covariance of X and Y: S (X, Y) = {(x ₁ −μ (x)) × (y ₁ −μ (y)) +.
+ (X _n −μ (x)) × (y _n −μ (y))} / n (5)

  The product moment correlation coefficient of X and Y is R (X, Y) = S (X, Y) / (σ (X) σ (Y)).

  For example, when the objective variable X = (5,6,9,4) and precipitation Y = (4,2,5,1) in the window W1, the average, variance, and standard deviation of X and Y are as follows: It is calculated as follows.

μ (X) = (5 + 6 + 9 + 4) / 4 = 24/4 = 6
μ (Y) = (4 + 2 + 5 + 1) / 4 = 2/4 = 3
σ ² (X) = ((5-6) ² + (6-6) ² + (9-6) ² + (4-6) ² ) / 4
= (1 + 0 + 9 + 4) /4=14/4=3.5
σ ² (Y) = ((4-3) ² + (2-3) ² + (5-3) ² + (1-3) ² ) / 4
= (1 + 1 + 4 + 4) /4=10/4=2.5
σ (X) = (3.5) ^1/2 1.81.87
σ (Y) = (2.5) ^{1/2 1.} 1.58

  Therefore, the covariance S (X, Y) of X and Y and the product moment correlation coefficient R (X, Y) are calculated as follows.

S (X, Y) = ((5-6) x (4-3) + (6-6) x (2-3) +
(9-6) × (5-3) + (4-6) × (1-3)) / 4
= (-1 + 0 + 6 + 4) /4=9/4=2.25
R (X, Y) ≒ 2.25 / (1.87 × 1.58) ≒ 2.25 / 2.95 ≒ 0.76

  When the correlation value is the absolute value of the product moment correlation coefficient, the correlation value is calculated as follows.

  Correlation value: | R (X, Y) | ≒ | 0.76 | = 0.76

  The classifying unit 42 classifies feature quantities having similar correlation changes calculated for each window into the same class, and classifies the feature quantities into a plurality of clusters. For example, for each feature quantity, the classification unit 42 obtains an error in correlation values with other feature quantities in each period. Then, the classifying unit 42 classifies feature quantities having similar errors into the same classification. FIG. 15 is a diagram illustrating an example of classifying data of a plurality of feature amounts. In the example of FIG. 15, the temperature and the electric energy are classified into cluster 1, the precipitation is classified into cluster 2, and the moving average value of the air temperature and the communication amount two units before the communication amount is classified into cluster 3. .

  For each cluster, the extraction unit 43 calculates a score by weighting the correlation value for each period calculated for the time-series data classified into the cluster with a predetermined weight. FIG. 16 is a diagram illustrating an example of a score for each cluster. In the example of FIG. 16, a weight is obtained by averaging the correlation values of the windows W <b> 1 to W <b> 3 with a weight having a correlation value of 0.7 or more being 1 and a weight having a correlation value of less than 0.7 being 0.

  The extraction unit 43 extracts a feature amount for each cluster based on the score. FIG. 17 is a diagram illustrating an example of feature amount extraction for each cluster. In the example of FIG. 17, the feature quantity having the maximum score is extracted for each cluster, the electric energy is extracted from the cluster 1, the precipitation is extracted from the cluster 2, and the moving average value of the traffic from the cluster 3 is extracted. Is extracted.

  As described above, the data output apparatus 10 can extract data having diversity in correlation with the objective variable, and can specify data effective for learning. Therefore, the prediction accuracy of machine learning can be improved.

[Flow of processing]
A flow of data output processing in which the data output apparatus 10 according to the first embodiment outputs diverse data will be described. FIG. 18 is a flowchart illustrating an example of the procedure of the data output process according to the first embodiment. The data output process is performed at a predetermined timing, for example, at a timing when an operation instruction to start the process is received from the input unit 21.

  As illustrated in FIG. 18, the calculation unit 41 divides the objective variable data stored in the objective variable data 30 and the plurality of feature amount data stored in the feature amount data 31 into windows (S10). The calculator 41 calculates, for each window, a correlation value between the value of the objective variable and a plurality of feature quantities (S11).

  The classification unit 42 classifies the feature quantities having similar changes in correlation calculated for each window into the same class, and classifies the feature quantities into a plurality of clusters (S12). The extraction unit 43 weights the correlation value of each feature amount for each window with a predetermined weight, and obtains a score from the weighted correlation value for each cluster (S13). The extraction unit 43 extracts the feature amount based on the score for each cluster (S14). The output unit 44 outputs information indicating the type of feature amount extracted from each of the plurality of clusters (S15), and ends the process.

[effect]
As described above, the data output device 10 according to the present embodiment has a time series data value indicating a specific monitoring target state and a value indicating a plurality of monitoring target states different from the specific monitoring target. Calculate the correlation value of each series data. The data output device 10 classifies time series data indicating the states of a plurality of monitoring targets into a plurality of clusters based on the correlation between the calculated correlation values. The data output device 10 extracts any time series data from each of the plurality of clusters. The data output device 10 outputs information indicating the type of each time-series data extracted from each of the plurality of clusters. Thereby, the data output device 10 can specify data effective for learning.

  Further, the data output apparatus 10 according to the present embodiment divides time-series data of values indicating a specific monitoring target state and time-series data of values indicating a plurality of monitoring target states into a plurality of periods. Do. The data output device 10 calculates, for each period, a correlation value between time-series data of values indicating a specific monitoring target state and time-series data of values indicating a plurality of monitoring target states. The data output device 10 classifies time-series data of values indicating a plurality of monitoring target states into a plurality of clusters based on the distribution of correlation values calculated for each period in a plurality of periods. Thereby, the data output device 10 can classify time-series data having similar data for each period into the same cluster.

  Further, the data output device 10 according to the present embodiment, for each cluster, based on a score obtained by weighting the correlation value for each period calculated for the time-series data classified into the cluster with a predetermined weight. Representative time-series data is extracted from time-series data classified into clusters. Thereby, the data output device 10 can extract representative time-series data having similar characteristics from each cluster.

  Next, Example 2 will be described. FIG. 19 is a diagram illustrating an example of a functional configuration of the data output apparatus according to the second embodiment. In addition, about the part similar to the data output device 10 which concerns on Example 1 shown in FIG. 1, the same code | symbol is attached | subjected and a different part is mainly demonstrated.

  The data output apparatus 10 according to the second embodiment outputs data effective for learning without classifying the feature quantities into clusters.

  The extraction unit 43A extracts time-series data based on the non-similarity of the correlation between the correlation values calculated by the calculation unit 41. For example, the extraction unit 43A extracts any time-series data from the time-series data indicating values of a plurality of monitoring targets based on the correlation between the correlation values, and the correlation value is similar to the extracted time-series data. The time series data is extracted by repeating excluding the time series data from the target of extraction. For example, the extraction unit 43A extracts one of the feature amounts, repeatedly removes the feature amount having a correlation value similar to the extracted feature amount from the extraction target, and has a variety of data in correlation with the objective variable. Extract

  The output unit 44 outputs information indicating the type of time-series data extracted by the extraction unit 43A. For example, the output unit 44 outputs information indicating the type of the feature amount extracted by the extraction unit 43A.

  This will be described using a specific example. FIG. 20 is a diagram illustrating an example of correlation values of feature amounts with respect to target variables. FIG. 20 shows the correlation values for the windows W1 to W3 of the precipitation, the temperature, the communication amount, the power amount, the temperature two unit time ago, and the moving average value of the communication amount. The extraction unit 43A calculates a score by weighting the correlation value of each window with predetermined weighting for each feature amount. In the example of FIG. 20, a weight having a correlation value of 0.7 or more is set to 1, a weight having a correlation value of less than 0.7 is set to 0, and a value obtained by weighted averaging the correlation values of the windows W1 to W3 is used as a score.

  The extraction unit 43A extracts any feature amount. For example, the extraction unit 43A extracts the feature amount having the highest score. FIG. 21 is a diagram illustrating an example in which the feature amount having the highest score is extracted. In the example of FIG. 21, precipitation is extracted.

  The extraction unit 43A excludes feature quantities having a correlation value similar to the extracted feature quantities from the extraction target. FIG. 22 is a diagram illustrating an example in which similar feature quantities are excluded from extraction targets. In the example of FIG. 22, feature quantities in which the Euclidean distance between the precipitation amount and the correlation value of each period is 0.5 or less are excluded.

  The extraction unit 43A extracts the feature amount having the highest score from the remaining feature amounts. FIG. 23 is a diagram illustrating an example in which the feature amount having the highest score is extracted from the remaining feature amounts. In the example of FIG. 23, the air temperature is extracted.

  The extraction unit 43A excludes feature quantities having a correlation value similar to the extracted feature quantities from the extraction target. FIG. 24 is a diagram illustrating an example in which similar feature quantities are excluded from extraction targets. In the example of FIG. 24, since there is no feature quantity with the Euclidean distance of the correlation value between the temperature and each period being 0.5 or less, none is excluded.

  The extraction unit 43A extracts the feature amount having the highest score from the remaining feature amounts. FIG. 25 is a diagram illustrating an example in which the feature amount having the highest score is extracted from the remaining feature amounts. In the example of FIG. 25, the temperature of 2 units time ago which remained is extracted.

  The output unit 44 outputs the amount of precipitation extracted by the extraction unit 43A, the temperature, and the temperature two units of time before.

[Flow of processing]
A flow of data output processing in which the data output device 10 according to the second embodiment outputs various data will be described. FIG. 26 is a flowchart illustrating an exemplary procedure of a data output process according to the second embodiment. The same parts as those of the data output process according to the first embodiment shown in FIG. 18 will be assigned the same reference numerals and descriptions thereof will be omitted.

  As illustrated in FIG. 26, the extraction unit 43A extracts any feature amount (S20). For example, the extraction unit 43A extracts the feature amount having the highest score. The extraction unit 43A excludes feature quantities whose correlation values are similar to the extracted feature quantities from the extraction targets (S21). The extraction unit 43A determines whether there is a feature amount to be extracted (S22). If there is a feature to be extracted (Yes at S22), the process proceeds to S20 described above, and any feature is extracted from the features to be extracted.

  On the other hand, when there is no feature quantity to be extracted (No at S22), the output unit 44 outputs information indicating the type of feature quantity extracted by the extraction unit 43A (S23), and ends the process.

[effect]
As described above, the data output device 10 according to the present embodiment has a time series data value indicating a specific monitoring target state and a value indicating a plurality of monitoring target states different from the specific monitoring target. Calculate the correlation value of each series data. The data output device 10 extracts time series data based on the dissimilarity of the correlation between the correlation values. The data output device 10 outputs information indicating the type of extracted time-series data. Thereby, the data output device 10 can specify data effective for learning.

  Further, the data output device 10 according to the present embodiment extracts any time series data from the time series data of values indicating the states of the plurality of monitoring targets based on the correlation between the calculated correlation values, and extracts the time series data. The time-series data is extracted by repeatedly removing time-series data having a correlation value similar to the extracted time-series data from the extraction target. As a result, the data output device 10 can extract time-series data having a variety of correlations with the specific monitoring target state.

  Although the embodiments of the disclosed apparatus have been described above, the disclosed technology may be implemented in various different forms other than the above-described embodiments.

  Further, each component of each illustrated apparatus is functionally conceptual, and does not necessarily need to be physically configured as illustrated. In other words, the specific state of distribution / integration of each device is not limited to the one shown in the figure, and all or a part thereof may be functionally or physically distributed or arbitrarily distributed in arbitrary units according to various loads or usage conditions. It can be integrated and configured. For example, the processing units of the reception unit 40, the calculation unit 41, the classification unit 42, the extraction unit 43 (extraction unit 43A), the output unit 44, and the prediction unit 45 may be integrated as appropriate. Further, the processing of each processing unit may be appropriately separated into a plurality of processing units. Further, all or any part of each processing function performed in each processing unit can be realized by a CPU and a program analyzed and executed by the CPU, or can be realized as hardware by wired logic. .

[Data output program]
The various processes described in the above embodiments can also be realized by executing a prepared program on a computer system such as a personal computer or a workstation. So, below, an example of a computer system which runs a program which has the same function as the above-mentioned example is explained. FIG. 27 shows a computer that executes a data output program.

  As shown in FIG. 27, the computer 300 includes a CPU (Central Processing Unit) 310, a HDD (Hard Disk Drive) 320, and a RAM (Random Access Memory) 340. These units 300 to 340 are connected via a bus 400.

  The HDD 320 stores in advance a data output program 320a that performs the same functions as the reception unit 40, the calculation unit 41, the classification unit 42, the extraction unit 43 (extraction unit 43A), the output unit 44, and the prediction unit 45. The data output program 320a may be separated as appropriate.

  The HDD 320 also stores various information. For example, the HDD 320 stores various data used for the OS and analysis.

  Then, the CPU 310 reads out and executes the data output program 320a from the HDD 320, thereby executing the same operation as each processing unit of the embodiment. That is, the data output program 320a performs the same operations as the reception unit 40, the calculation unit 41, the classification unit 42, the extraction unit 43 (extraction unit 43A), the output unit 44, and the prediction unit 45.

  Note that the data output program 320a described above does not necessarily have to be stored in the HDD 320 from the beginning.

  For example, the program is stored in a “portable physical medium” such as a flexible disk (FD), a CD-ROM, a DVD disk, a magneto-optical disk, or an IC card inserted into the computer 300. Then, the computer 300 may read and execute programs from these.

  Furthermore, the program is stored in “another computer (or server)” connected to the computer 300 via a public line, the Internet, a LAN, a WAN or the like. Then, the computer 300 may read and execute programs from these.

10 Data Output Device 23 Storage Unit 24 Control Unit 30 Objective Variable Data 31 Feature Quantity Data 40 Reception Unit 41 Calculation Unit 42 Classification Unit 43, 43A Extraction Unit 44 Output Unit 45 Prediction Unit

Claims (3)

Each of time-series data of a value indicating a state of a specific monitoring target and time-series data of a value indicating a state of a plurality of monitoring targets different from the specific monitoring target is divided into a plurality of periods;
For each of the periods , correlation values between time series data of values indicating the status of the specific monitoring target and time series data of values indicating the status of the plurality of monitoring targets are calculated;
The time series data indicating the states of the plurality of monitoring targets are classified into a plurality of clusters based on the distribution of the correlation value calculated for each period in the plurality of periods ,
For each of the time-series data indicating the state of the plurality of monitoring targets, a score is obtained by performing a weighting operation with a predetermined weight on the correlation value calculated for the time-series data, and for each cluster, Extract time-series data with the maximum score from the time-series data classified into the cluster ,
Outputting information indicating the type of each time-series data extracted from each of the plurality of clusters;
And a computer executes the processing . Each of time-series data of a value indicating a state of a specific monitoring target and time-series data of a value indicating a state of a plurality of monitoring targets different from the specific monitoring target is divided into a plurality of periods;
For each of the periods , correlation values between time series data of values indicating the status of the specific monitoring target and time series data of values indicating the status of the plurality of monitoring targets are calculated;
The time series data indicating the states of the plurality of monitoring targets are classified into a plurality of clusters based on the distribution of the correlation value calculated for each period in the plurality of periods ,
For each of the time-series data indicating the state of the plurality of monitoring targets, a score is obtained by performing a weighting operation with a predetermined weight on the correlation value calculated for the time-series data, and for each cluster, Extract time-series data with the maximum score from the time-series data classified into the cluster ,
Outputting information indicating the type of each time-series data extracted from each of the plurality of clusters;
A data output program that causes a computer to execute a process. Each of time-series data of a value indicating a state of a specific monitoring target and time-series data of values indicating a state of a plurality of monitoring targets different from the specific monitoring target is divided into a plurality of periods, and each period A calculation unit that calculates correlation values between time-series data of values indicating the status of the specific monitoring target and time-series data of values indicating the status of the plurality of monitoring targets ;
A classification unit that classifies time series data indicating the states of the plurality of monitoring targets into a plurality of clusters based on the distribution of the correlation value calculated for each period by the calculation unit in the plurality of periods ;
For each of the time-series data indicating the state of the plurality of monitoring targets, a score is obtained by performing a weighting operation with a predetermined weight on the correlation value calculated for the time-series data, and for each cluster, An extraction unit for extracting time-series data having the largest score from time-series data classified into the cluster ;
An output unit that outputs information indicating a type of each time-series data extracted from each of the plurality of clusters by the extraction unit;
A data output device comprising:

Images