Time-series data processing device, time-series data processing method, and program

Abstract

A time-series data processing device (1) comprises: an acquisition unit (11) that acquires metadata, including text information related to an objective variable and attribute information related to an explanatory variable candidate, and time-series data of the objective variable and the explanatory variable candidate; and a relevance determination unit (12) that uses the metadata and the time-series data to determine a function that defines the relationship between the objective variable and the explanatory variable candidate.

Description

Time series data processing device, time series data processing method and program

[0001] The present disclosure relates to a time series data processing device, a time series data processing method, and a program.

[0002] Efforts are being made to analyze the temporal fluctuation trends of data and utilize the analysis results in various fields. For example, Patent Literature 1 describes an information processing device that selects candidate explanatory variables from data having an absolute value of a correlation coefficient equal to or greater than a predetermined value with respect to data corresponding to a target variable.

[0003] International Application No. 2017 / 203601

[0004] In the conventional technology described in Patent Document 1, the relationship between a dependent variable and an explanatory variable is defined only by a correlation coefficient that indicates the statistical relationship between the time-series data of the dependent variable and the time-series data of the explanatory variable. Therefore, there is a problem in that an explanatory variable that has a semantic relationship with the dependent variable may not be selected as an explanatory variable to be used for predicting the dependent variable.

[0005] The present disclosure is intended to solve the above-mentioned problems, and aims to provide a time-series data processing device that can select explanatory variables that have a statistical correlation with time-series data of a dependent variable and a semantic relationship with the dependent variable as explanatory variables to be used in predicting the dependent variable.

[0006] A time-series data processing device according to the present disclosure includes an acquisition unit that acquires metadata including text information on a dependent variable and attribute information on candidate explanatory variables, and time-series data on the dependent variable and the candidate explanatory variables, and an association determination unit that determines a function that defines the relationship between the dependent variable and the candidate explanatory variables using the metadata and the time-series data.

[0007] According to the present disclosure, a function defining the relationship between a dependent variable and candidate explanatory variables is determined using metadata and time-series data. The time-series data includes a statistical correlation between the dependent variable and the explanatory variables, and the metadata includes a semantic correlation between the dependent variable and the explanatory variables. This allows the time-series data processing device according to the present disclosure to select explanatory variables that are statistically correlated with the time-series data of the dependent variable and semantically related to the dependent variable as explanatory variables to be used in predicting the dependent variable.

[0008] 1 is a block diagram showing an example of a configuration of a time series data processing device according to embodiment 1. FIG. 3A and FIG. 3B are diagrams showing examples of keywords related to a dependent variable and attribute information related to explanatory variable candidates. FIG. 4A, FIG. 4B, and FIG. 4C are diagrams showing examples of attribute information related to explanatory variable candidates, keywords related to the dependent variable, and semantic relevance. FIG. 5A and FIG. 5B are block diagrams showing a hardware configuration for realizing the functions of the time series data processing device according to embodiment 1. FIG. 5A is a block diagram showing an example of a configuration of a time series data processing device according to embodiment 2. FIG. 6A is a flowchart showing a time series data processing method according to embodiment 6. FIG. 6B is a diagram showing examples of attribute information, attribute pairs, hierarchical weights, and semantic similarities related to explanatory variable candidates. FIG. 7A is a diagram showing a similarity matrix and distances between explanatory variables. FIG. 7B is a diagram showing a similarity matrix and distances between explanatory variables. FIG. 7C is a diagram showing an example of a configuration of a time series data processing device according to embodiment 7. FIG. 7D is a diagram showing an example of a screen (1) in embodiment 3. FIG. 7E is a diagram showing an example of a screen (2) in embodiment 3. FIG. 7F is a diagram showing an example of a screen (3) in embodiment 3. FIG. 7F is a block diagram showing an example of a configuration of a time series data processing device according to embodiment 7. 13 is a flowchart showing the operation of a testing unit in the fourth embodiment.

[0009] First Embodiment A time-series data processing device according to a first embodiment selects, from among explanatory variable candidates, explanatory variables that are statistically and semantically related to a target variable.

[0010] (Basic Configuration of a Time Series Data Processing Device) FIG. 1 is a block diagram showing an example configuration of a time series data processing device 1 according to embodiment 1. The time series data processing device 1 selects explanatory variables to be used for predicting a target variable from among multiple types of explanatory variables. The explanatory variables are time series data and are parameters detected in a time series. For example, the explanatory variables are parameters that change over time. The way in which the explanatory variables change over time may be arbitrary.

[0011] The dependent variable is a parameter that is affected by multiple types of explanatory variables, and its value changes when the values ​​of the explanatory variables change. The dependent variable is also time-series data whose value changes depending on the explanatory variables and over time. For example, if the dependent variable is the demand for air conditioners, the explanatory variables could be data on sales of air conditioners over a certain period of time.

[0012] 1, the time series data processing device 1 includes an acquisition unit 11, a relevance determination unit 12, and an extraction unit 13. For example, the time series data processing device 1 is realized by a computer. A memory included in the computer stores a program constituting an information processing application for realizing each function of the acquisition unit 11, the relevance determination unit 12, and the extraction unit 13. A processor included in the computer executes the information processing application read from the memory, thereby realizing each function of the acquisition unit 11, the relevance determination unit 12, and the extraction unit 13.

[0013] (Acquisition Unit) The acquisition unit 11 acquires time-series data and metadata. For example, it is assumed that the time-series data processing device 1 includes a communication unit (not shown in FIG. 1 ) and that this communication unit enables communication connection with an external device (not shown in FIG. 1 ). In this case, the acquisition unit 11 controls the communication unit to acquire time-series data and metadata from the external device. The external device is, for example, a data server. Examples of communication methods used for communication with the external device include Bluetooth (registered trademark), Wi-Fi, Zigbee (registered trademark), LoRa (registered trademark), and NFC (Near Field Communication). The communication unit is, for example, a wireless communication module compatible with any of these communication methods.

[0014] The acquisition unit 11 also controls the communication unit to execute procedures such as pairing, network connection, and address assignment to establish communication with the external device. Once communication is established, the communication unit requests time-series data and metadata from the external device, and the external device returns the requested data. The data received by the communication unit is acquired by the acquisition unit 11.

[0015] The time series data is time series data related to the dependent variable and time series data related to the candidate explanatory variables. Identification information for identifying the data is assigned to the time series data. For example, as identification information, the time series data related to the dependent variable is assigned the data name "demand for air conditioner A," and the time series data related to the candidate explanatory variables is assigned the data name "data related to sales of air conditioner A in period B." Hereinafter, the names of the time series data of multiple candidate explanatory variables will be referred to as "explanatory variable a, explanatory variable b, explanatory variable c, ...." Furthermore, the sampling periods of the time series data of the dependent variable and the candidate explanatory variables are all the same. For example, all data may be monthly data or all data may be daily data.

[0016] Metadata is data that includes keywords related to the objective variable and attribute information related to the explanatory variable candidates. Keywords are text information related to the objective variable and are specified by the user. For example, if the objective variable is "demand for a certain product," keywords are determined based on information such as the product's use, destination, related parts, and leading indicators. More specifically, if the objective variable is "demand for air conditioner products," there are keywords that indicate the air conditioner products themselves, such as "air conditioner," "air conditioner," and "outdoor unit."

[0017] The attribute information on the explanatory variable candidates is attribute or category information that indicates the classification of data assigned to each explanatory variable candidate. The attribute information may be structured in multiple layers. For example, if the objective variable is "demand for air conditioner products," a keyword related to this objective variable is "air conditioner." The attribute information on the explanatory variable candidates that affect this objective variable includes the major classification information "machinery" because "air conditioner" is a machine, and the medium classification information "inventory" related to demand.

[0018] (Relevance Degree Determination Unit) The relevance degree determination unit 12 determines a function that defines the relationship between the objective variable and the explanatory variable candidates using metadata and time-series data. For example, the relevance degree determination unit 12 determines a function type according to the objective variable and the explanatory variable candidates from among various function types that calculate numerical information about the relationship between the objective variable and the explanatory variable candidates using metadata and time-series data. The function type may be any type that calculates numerical information when metadata and time-series data are substituted.

[0019] Determining the function also includes selecting a trained model generated by machine learning. For example, when metadata and time-series data are input, the relevance determination unit 12 selects a trained model that outputs numerical information indicating the relationship between the objective variable and candidate explanatory variables. The trained model may be configured using an algorithm such as Transformer, BERT (Bidirectional Encoder Representations from Transformers), or GPT (Generative Pre-Training), or may be configured by combining multiple algorithms including these.

[0020] (Extraction Unit) The extraction unit 13 extracts explanatory variables from the explanatory variable candidates based on a comparison between the function value determined by the relevance determination unit 12 and a threshold. For example, the extraction unit 13 extracts explanatory variable candidates corresponding to function values ​​exceeding a threshold from among multiple explanatory variable candidates. This enables the time series data processing device 1 to extract explanatory variables effective for predicting the dependent variable. Note that the threshold is a threshold related to the function value that is preset for the extraction unit 13, and is determined based on, for example, a prediction result of the dependent variable using the extracted explanatory variables.

[0021] 1 , the extraction unit 13 is included in the time series data processing device 1, but it may be included in an external device separate from the time series data processing device 1. For example, the relevance determination unit 12 connects to the external device via a communication unit included in the time series data processing device 1 and transmits information indicating the determined function to the external device. The extraction unit included in the external device extracts explanatory variables from the explanatory variable candidates based on a comparison between the value of the function received from the time series data processing device 1 and a threshold. For example, the external device may be a prediction device that predicts a target variable based on the extracted explanatory variables.

[0022] Next, a time-series data processing method according to the first embodiment will be described. FIG. 2 is a flowchart showing the time-series data processing method according to the first embodiment, illustrating a series of processes performed by the time-series data processing device 1. The acquisition unit 11 acquires time-series data related to the dependent variable, time-series data related to candidate explanatory variables, and metadata including keywords related to the dependent variable and attribute information related to the candidate explanatory variables (step ST1). For example, the time-series data related to the dependent variable and the time-series data related to the candidate explanatory variables are data sets acquired at regular intervals.

[0023] The dependent variable is data that is the target of prediction or analysis. When the dependent variable is "demand for air conditioner products," the time-series data related to the dependent variable is, for example, time-series data on the demand for air conditioner products over a certain period of time. The explanatory variable is data that can affect the dependent variable. When the dependent variable is "demand for air conditioner products," candidate explanatory variables that affect the demand include economic statistics data, meteorological data, and product data.

[0024] When the candidate explanatory variable is "economic statistical data," the time series data may include data on orders, production, shipments, sales, and inventory for electronic components, electrical equipment, machinery, etc., or data on construction statistics, demographics, the labor market, foreign trade, banking statistics, energy, etc. When the candidate explanatory variable is "weather data," the time series data may include temperature (daily maximum temperature, minimum temperature, average temperature, etc.), humidity (daily average humidity, maximum humidity, minimum humidity, etc.), precipitation (daily precipitation, etc.), wind speed (average wind speed, maximum wind speed), etc. When the candidate explanatory variable is "product data," the time series data may include inventory levels (current inventory quantities), price fluctuations (fluctuations in product prices), etc.

[0025] The metadata includes keywords as meta information related to the dependent variable and attribute information as meta information related to the candidate explanatory variables. For example, assume that metadata related to the target of prediction or analysis is generated in advance and stored in a data server. FIG. 3A is a diagram showing keywords related to the dependent variable. In FIG. 3A, the dependent variable is the demand for air conditioner products. In this case, as shown in FIG. 3A, keywords related to the dependent variable include "air conditioner," "air conditioner," and "outdoor unit," which are related to air conditioner products. Furthermore, based on "start of construction," which can be a leading indicator of the flow of demand generation for air conditioner products, keywords such as "construction," "start of construction," "building," and "housing" are listed.

[0026] FIG. 3B illustrates attribute information related to explanatory variable candidates. For example, in FIG. 3B, explanatory variables a, b, c, ..., h, i, ... are explanatory variable candidates. Explanatory variables a, b, and c correspond to the objective variable whose keywords are "construction" and "construction start." Explanatory variables a and b contain information related to "construction" and "construction start." In the attribute information for explanatory variables a and b, the major category is "construction" and the intermediate category is "construction start." Explanatory variable c contains information related to "construction" and "orders." In the attribute information for explanatory variable c, the major category is "construction" and the intermediate category is "orders." While FIG. 3B illustrates a case in which explanatory variables correspond to keywords, the two do not necessarily need to correspond to each other. For example, the time series data processing device 1 can determine whether to remove explanatory variables that are unrelated to the objective variable and do not correspond to keywords by threshold processing based on their relevance.

[0027] The explanatory variables h and i correspond to the objective variables whose keywords are "air conditioner," "air conditioner," and "outdoor unit." The explanatory variable h includes information belonging to "machinery" corresponding to "air conditioner," "air conditioner," or "outdoor unit," etc., and includes information about its "production." As shown in FIG. 3B , the attribute information of the explanatory variable h includes the major category of "machinery" and the medium category of "production." Similarly, the explanatory variable i includes information belonging to "machinery" corresponding to "air conditioner," "air conditioner," or "outdoor unit," etc., and includes information about its "inventory." The attribute information of the explanatory variable i includes the major category of "machinery" and the medium category of "inventory."

[0028] As the relevance determination process, the relevance determiner 12 determines a function that defines the relationship between the dependent variable and the explanatory variable candidates using the metadata and the time-series data (step ST2). For example, the relevance determiner 12 determines a function for calculating a degree of relevance for extraction that is a combination of the statistical relevance of the time-series data between the dependent variable and the explanatory variable candidates and the semantic relevance between the dependent variable and the explanatory variable candidates based on the metadata. This enables the time-series data processing device 1 to select explanatory variables that are semantically related to the dependent variable in addition to the statistical relevance with the time-series data of the dependent variable.

[0029] The statistical relevance is an index that indicates the statistical closeness between the time series data for the dependent variable and the time series data for the candidate explanatory variables. This relevance is the degree of relevance that takes into account the data trend, small fluctuations, or phase shift between the waveforms formed by the time series data for the dependent variable and the waveforms formed by the time series data for the candidate explanatory variables.

[0030] Semantic relevance is an index showing the semantic closeness between keywords related to the objective variable and attribute information related to the explanatory variable candidates. The attribute information related to the explanatory variable candidates is related to the specific field or area to which the explanatory variable candidates belong, and can also be considered domain knowledge. Therefore, semantic relevance is also an index of the domain knowledge relevance between keywords related to the objective variable and attribute information related to the explanatory variable candidates.

[0031] (Preprocessing) The relevance determination unit 12 performs missing value processing, standardization processing, and data conversion processing on the time-series data of the dependent variable and the explanatory variable candidates, and calculates the value of a function using the processed time-series data. For example, the relevance determination unit 12 performs the above processing on the time-series data related to the dependent variable and the time-series data related to the explanatory variable candidates acquired by the acquisition unit 11. This makes it possible to more accurately calculate the relationship between the dependent variable and the explanatory variable candidates using the function determined by the relevance determination unit 12.

[0032] Missing value processing is a process of interpolating missing values ​​in time series data, and includes, for example, processes that interpolate using previous and previous values ​​in time (carry forward method, backward fill method), linear interpolation, and time-weighted interpolation. The carry forward method interpolates missing values ​​using previous values ​​in time, and the backward fill method interpolates missing values ​​using later values ​​in time. Linear interpolation is a process that linearly interpolates missing values ​​using previous and previous data in time. Time-weighted interpolation is a process that interpolates using the average or median value of the data over a specific period.

[0033] Standardization is a process of unifying the scale of data to facilitate comparison between different feature quantities, and examples of this include standardization and normalization. Standardization is a process of setting the mean of data to 0 and the standard deviation to 1. Normalization is a process of scaling data to a range from 0 to 1.

[0034] Data transformation processing is processing that changes the properties of data to make analysis or modeling easier, and includes, for example, moving average, difference series, logarithmic transformation, seasonal adjustment, or a combination of these. Moving average is processing that divides the sum of data in a specified range by the number of data items in order to smooth the data. Difference series is processing that takes the difference from the previous value in time to remove trends. Logarithmic transformation is processing that converts data into logarithms when the data distribution is close to a log-normal distribution. Seasonal adjustment is processing that adjusts data at a specific cycle to remove seasonality. By performing these processes or a combination of these processes on time series data, the relevance determination unit 12 can calculate more accurate relevance for extraction.

[0035] The relevance determination unit 12 calculates the extraction relevance as a function value for the determined function using the processed time-series data. The extraction relevance calculated by the relevance determination unit 12 is output to the extraction unit 13.

[0036] 2 , the extraction unit 13 extracts explanatory variables from the explanatory variable candidates based on a comparison between the relevance for extraction and a threshold (step ST3). For example, the extraction unit 13 extracts explanatory variables corresponding to the relevance for extraction that exceed the threshold from among the multiple explanatory variable candidates. Hereinafter, the explanatory variables extracted by the extraction unit 13 will be referred to as first extracted explanatory variable data. For example, the first extracted explanatory variable data extracted by the time series data processing device 1 is output to a prediction device. The prediction device predicts the dependent variable using the first extracted explanatory variable data. By performing the processes from step ST1 to step ST3, the time series data processing device 1 can select explanatory variables that are statistically related to the time series data of the dependent variable and semantically related to the dependent variable as the first extracted explanatory variable data to be used for predicting the dependent variable.

[0037] (Specific Example of Relevance Calculation) The function for calculating the relevance for extraction is a function having a first part and a second part. The first part is a part for calculating a correlation coefficient, dynamic time warping (hereinafter referred to as DTW), Euclidean distance, or a combination of these, for the time-series data between the dependent variable and the explanatory variable candidates, as the statistical relevance of the time-series data between the dependent variable and the explanatory variable candidates. The second part is a part for calculating the cosine similarity between the vectorized text information of the dependent variable and the attribute information of the explanatory variable candidates, as the semantic relevance between the text information of the dependent variable and the attribute information of the explanatory variable candidates.

[0038] (Calculation of Statistical Relevance) In the first part, the statistical relevance of the time series data is calculated for each candidate explanatory variable. For example, the first part is composed of a combination of three indices for calculating the statistical relevance: correlation coefficient, DTW, and Euclidean distance. In the first part, the score obtained for each indices is converted so that the value ranges from 0 (low) to 1 (high), and the statistical relevance is calculated by taking a weighted average of these scores.

[0039] The time series data (x, y) of the objective variable and the candidate explanatory variables input to the first part is normalized by the relevance determination unit 12 so that the value of each point falls within the range of 0 to 1 according to the following formula (1): x'=(x-min(x)) / (max(x)-min(x)) y'=(y-min(y)) / (max(y)-min(y)) (1)

[0040] The correlation coefficient, which is an index for calculating the statistical relevance, is a statistic that indicates the strength of the linear relationship between the time series data related to the objective variable and the time series data related to the candidate explanatory variables. The value of the correlation coefficient ranges from -1 to 1, and the closer to 1, the stronger the positive correlation, and the closer to -1, the stronger the negative correlation. The relevance determination unit 12 calculates the correlation coefficient s for a pair of normalized data x', y' according to the following formula (2): ｃｏｒ Calculate s ｃｏｒ = | cor (x', y') | ∈ [0, 1] (2)

[0041] DTW is an index for evaluating the similarity of time series data, and is effective even when the time axes are misaligned. DTW calculates the distance by optimally aligning two time series data. The relevance determination unit 12 calculates a DTW score s for a pair of normalized data x', y', taking into account positive and negative values ​​according to the following formula (3): ｄｔｗ In the following formula (3), N is the length of the time series data, and dtw(x', y') is the DTW distance. ｄｔｗ ｐ is the score of dtw(x', y'), which is the normal DTW distance. ｄｔｗ ｎ is the score when the data is inverted. ｄｔｗ is the final score, a value between 0 and 1. ｄｔｗ ｐ =1-(dtw(x',y')) / Ns ｄｔｗ ｎ =1-(dtw(1-x',y')) / Ns ｄｔｗ = max(s ｄｔｗ ｐ , s ｄｔｗ ｎ ) ∈ [0, 1] (3)

[0042] The Euclidean distance is the linear distance between two data points, and in the case of time-series data, it is calculated by comparing values ​​at the same time point. The relevance determination unit 12 calculates the score s of the Euclidean distance for the pair x', y' of normalized data, taking into account positive and negative values, according to the following formula (4): Ｌ２ In the following formula (4), N is the length of the time series data. ２ (x', y') is the Euclidean distance. Ｌ２ ｐ is the normal Euclidean distance L ２ is the score of (x', y'). Ｌ２ ｎ is the score when the data is inverted. Ｌ２ is the final score, a value between 0 and 1. Ｌ２ ｐ = 1 - (L ２ (x', y') / √N) s Ｌ２ ｎ = 1 - (L ２ (1-x', y') / √N) s Ｌ２ = max(s Ｌ２ ｐ , s Ｌ２ ｎ ) ∈ [0, 1] (4)

[0043] The relevance determination unit 12 calculates the weighted average value s of the above three indices as the statistical relevance in accordance with the following formula (5): In the following formula (5), i is a subscript indicating the three indices, namely, the correlation coefficient, the DTW, and the Euclidean distance. ｉ is the score of the index indicated by the subscript i, and the correlation coefficient score s ｃｏｒ , DTW score s ｄｔｗ , and the Euclidean distance score s Ｌ２ That is. ｉ is the weight given to the score of each index. ｉ∈｛ｄｔｗ，ｃｏｒ，Ｌ２｝ w ｉ s ｉ ∈[0, 1] Σ ｉ∈｛ｄｔｗ，ｃｏｒ，Ｌ２｝ w ｉ = 1 (5)

[0044] (Calculation of Semantic Relevance) In the second part, a semantic relevance is calculated by quantifying the degree of match between the keyword related to the objective variable designated by the user and the attribute information related to the explanatory variable candidates.

[0045] For example, in the second part, the keywords of the objective variable and the attribute information of the candidate explanatory variables are vectorized, and the cosine similarity between the keyword vector and the attribute information vector is calculated as the semantic relevance. Vectorization (numerical representation) is achieved by using word2vec, BERT, or the like. word2vec is a process of mapping words to fixed-length vectors. BERT outputs vector information when tokens converted from text information are input, and the BERT model outputs vector information when tokens are input. These vectors can capture the semantic relationship between words. However, if the cosine similarity is a negative value, the vector is set to 0.

[0046] FIG. 4A is a diagram showing an example of attribute information related to explanatory variable candidates. In FIG. 4A, explanatory variable a, explanatory variable b, explanatory variable c, ..., explanatory variable h, explanatory variable i, ... are explanatory variable candidates. Explanatory variables a, b, and c correspond to objective variables whose keywords are "construction" and "start of construction." Explanatory variables a and b include information belonging to "construction" and "start of construction," etc. In the attribute information of explanatory variables a and b, the major category is "construction" and the medium category is "start of construction." Explanatory variable c includes information belonging to "construction" and "order," etc. In the attribute information of explanatory variable c, the major category is "construction" and the medium category is "order."

[0047] The explanatory variables h and i correspond to the objective variables whose keywords are "air conditioner," "air conditioner," and "outdoor unit." The explanatory variable h includes information belonging to "machinery" corresponding to "air conditioner," "air conditioner," or "outdoor unit," etc., and includes information related to its "production." As shown in FIG. 4A , the attribute information of the explanatory variable h includes the major category of "machinery" and the medium category of "production." Similarly, the explanatory variable i includes information belonging to "machinery" corresponding to "air conditioner," "air conditioner," or "outdoor unit," etc., and includes information related to its "inventory." The attribute information of the explanatory variable i includes the major category of "machinery" and the medium category of "inventory."

[0048] FIG. 4B is a diagram showing an example of keywords related to the objective variable. In FIG. 4B, the objective variable is the demand for air conditioner products. In this case, as shown in FIG. 4B, the keywords related to the objective variable include "air conditioner," "air conditioner," and "outdoor unit" related to air conditioner products. Furthermore, based on "construction start," which can be a leading indicator of the flow of demand generation for air conditioner products, "construction," "construction start," "building," or "housing" are listed as keywords. In the example shown in FIG. 4B, weights are assigned to keywords related to the objective variable according to their specificity. Alternatively, weights may be assigned in descending order of priority. A weight of "1.00" is assigned to the specific keywords "air conditioner" and "construction," while a weight of "0.95" is assigned to the more abstract keyword "construction start."

[0049] FIG. 4C is a diagram showing an example of semantic relevance (semantic relevance). In FIG. 4C, explanatory variable a, explanatory variable b, explanatory variable c, ..., explanatory variable h, explanatory variable i, ... are the explanatory variable candidates shown in FIG. 4A. As shown in FIG. 4B, when there are multiple keywords for a certain objective variable, the relevance determination unit 12 calculates the semantic relevance for each keyword. For example, the explanatory variable candidates corresponding to "architecture," a keyword related to the objective variable, are explanatory variables a, b, and c, whose attribute information falls under the major category of "architecture." Furthermore, the explanatory variable candidates corresponding to "air conditioner," a keyword related to the objective variable, are explanatory variables h and i, whose attribute information falls under the major category of "machine."

[0050] When there are multiple keywords, the relevance determination unit 12 calculates the semantic relevance for each of these keywords (process (1)). For example, when there are five keywords, the relevance determination unit 12 calculates five semantic relevances corresponding to the respective keywords for the explanatory variable a.

[0051] If a weight for specificity or priority is assigned to a keyword, the relevance determination unit 12 multiplies the semantic relevance by the weight (process (2)). For example, as shown in FIG. 4B, if a weight of 0.95 is assigned to the keyword "construction commenced," the semantic relevance corresponding to the keyword "construction commenced" for the explanatory variable a is multiplied by the weight of 0.95. As a result, the semantic relevance corresponding to a keyword with high specificity or high priority becomes higher than the semantic relevance of a keyword with low specificity or low priority. Note that if a weight for specificity or priority is not assigned, process (2) of multiplying the semantic relevance by the weight is omitted.

[0052] Then, the relevance determination unit 12 adopts the largest value of the calculated semantic relevance (process (3)). For example, if there are five keywords, the largest value of the five semantic relevance is adopted.

[0053] If the attribute information regarding the explanatory variable candidates is hierarchical information, the semantic relevance is calculated for each layer. For example, if there are multiple layers, such as the "major categories" and "middle categories" shown in FIG. 4A, the relevance determination unit 12 performs processes (1), (2), and (3) for each layer to calculate the semantic relevance for each layer. The relevance determination unit 12 calculates a weighted average of the semantic relevance for each layer and obtains one semantic relevance for each explanatory variable. For example, if there are two layers, two semantic relevances are calculated for each explanatory variable, and a weighted average of the two semantic relevances is calculated. Note that the weights may be specified by the user. For example, the weights may be increased in order of the priority set for the keyword or the specificity.

[0054] For example, in the function determined by the relevance determiner 12, a weighted average of the statistical relevance calculated by the first part and the semantic relevance calculated by the second part is calculated as the relevance for extraction. This enables the time-series data processing device 1 to select explanatory variables that have a statistical relevance with the time-series data of the dependent variable and a semantic relationship with the dependent variable as the first extracted explanatory variable data to be used for predicting the dependent variable.

[0055] Next, a description will be given of the hardware configuration that realizes the functions of the time series data processing device 1. The functions of the acquisition unit 11, relevance determination unit 12, and extraction unit 13 included in the time series data processing device 1 are realized by processing circuits. That is, the time series data processing device 1 includes a processing circuit for executing the processes from step ST1 to step ST3 shown in Fig. 2. The processing circuit may be dedicated hardware, or may be a CPU (Central Processing Unit) that executes a program stored in memory.

[0056] FIG. 5A is a block diagram showing a hardware configuration for realizing the functions of the time-series data processing device 1. FIG. 5B is a block diagram showing a hardware configuration for executing software for realizing the functions of the time-series data processing device 1. In FIGS. 5A and 5B , the acquisition unit 11 acquires time-series data related to the dependent variable, time-series data related to candidate explanatory variables, keywords related to the dependent variable, and attribute information related to the candidate explanatory variables from an external device (e.g., a data server) via the input interface 100. The extraction unit 13 outputs the extracted explanatory variables to an external device (e.g., a prediction device) via the output interface 101.

[0057] 5A, the processing circuit 102 may be, for example, a single circuit, a composite circuit, a programmed processor, a parallel programmed processor, an ASIC (Application Specific Integrated Circuit), an FPGA (Field-Programmable Gate Array), or a combination thereof. The functions of the acquisition unit 11, the relevance determination unit 12, and the extraction unit 13 included in the time-series data processing device 1 may be realized by separate processing circuits, or these functions may be realized together by a single processing circuit.

[0058] 5B , the functions of the acquisition unit 11, the relevance determination unit 12, and the extraction unit 13 included in the time-series data processing device 1 are realized by software, firmware, or a combination of software and firmware. The software or firmware is written as a program and stored in the memory 104.

[0059] The processor 103 reads and executes programs stored in the memory 104 to realize the functions of the acquisition unit 11, the relevance determination unit 12, and the extraction unit 13 included in the time series data processing device 1. For example, the time series data processing device 1 includes a memory 104 for storing a program that, when executed by the processor 103, results in the execution of steps ST1 to ST3 shown in FIG. 2 . These programs cause a computer to execute the procedures or methods of the processes performed by the acquisition unit 11, the relevance determination unit 12, and the extraction unit 13. The memory 104 may be a computer-readable storage medium that stores programs for causing a computer to function as the acquisition unit 11, the relevance determination unit 12, and the extraction unit 13.

[0060] The memory 104 may be, for example, a non-volatile or volatile semiconductor memory such as a RAM (Random Access Memory), a ROM (Read Only Memory), a flash memory, an EPROM (Erasable Programmable Read Only Memory), or an EEPROM (Electrically-EPROM) (registered trademark), a magnetic disk, a flexible disk, an optical disk, a compact disk, a mini disk, or a DVD.

[0061] Some of the functions of the acquisition unit 11, the relevance determination unit 12, and the extraction unit 13 included in the time-series data processing device 1 may be realized by dedicated hardware, and the other functions may be realized by software or firmware. For example, the function of the acquisition unit 11 may be realized by a processing circuit 102 that is dedicated hardware, and the functions of the relevance determination unit 12 and the extraction unit 13 may be realized by the processor 103 reading and executing a program stored in the memory 104. In this way, the processing circuit can realize the above functions by hardware, software, firmware, or a combination of these.

[0062] As described above, the time series data processing device 1 according to the first embodiment includes an acquisition unit 11 that acquires metadata including text information on a dependent variable and attribute information on explanatory variable candidates, and time series data on the dependent variable and the explanatory variable candidates, and an association determination unit 12 that determines a function defining the relationship between the dependent variable and the explanatory variable candidates using the metadata and the time series data. The time series data includes a statistical association between the dependent variable and the explanatory variables, and the metadata includes a semantic association between the dependent variable and the explanatory variables. This allows the time series data processing device 1 to select explanatory variables that are statistically associated with the time series data of the dependent variable and semantically related to the dependent variable as explanatory variables to be used for predicting the dependent variable.

[0063] The time-series data processing device 1 according to the first embodiment includes an extracting unit 13 that extracts first extracted explanatory variable data from the explanatory variable candidates based on a comparison between the value of the function determined by the relevance determining unit 12 and a threshold value. This enables the time-series data processing device 1 to extract explanatory variables that are effective for predicting the dependent variable.

[0064] In the time-series data processing device 1 according to the first embodiment, the relevance determiner 12 determines a function for calculating a degree of relevance for extraction, which is a combination of the statistical relevance of the time-series data between the dependent variable and the explanatory variable candidates and the semantic relevance between the dependent variable and the explanatory variable candidates based on the metadata. The extractor 13 extracts first extracted explanatory variable data from the explanatory variable candidates based on a comparison between the function-extraction relevance determined by the relevance determiner 12 and a threshold. This enables the time-series data processing device 1 to select explanatory variables that have a statistical relevance with the time-series data of the dependent variable and a semantic relationship with the dependent variable, as explanatory variables to be used for predicting the dependent variable.

[0065] In the time-series data processing device 1 according to the first embodiment, the function determined by the relevance determiner 12 includes a portion for calculating a correlation coefficient, dynamic time warping, or Euclidean distance, or a combination of these, between the time-series data of the dependent variable and the explanatory variable candidates as the statistical relevance of the time-series data between the dependent variable and the explanatory variable candidates, and a portion for calculating a cosine similarity between the vectorized text information of the dependent variable and the attribute information of the explanatory variable candidates as the semantic relevance between the text information of the dependent variable and the attribute information of the explanatory variable candidates. This enables the time-series data processing device 1 to select explanatory variables that are statistically relevant to the time-series data of the dependent variable and that are semantically related to the dependent variable, as explanatory variables to be used for predicting the dependent variable.

[0066] In the time-series data processing device 1 according to the first embodiment, the relevance determiner 12 performs missing value processing, standardization processing, and data conversion processing on the time-series data of the dependent variable and the explanatory variable candidates, and calculates the value of the function using the processed time-series data. This enables the time-series data processing device 1 to more accurately calculate the relationship between the dependent variable and the explanatory variable candidates using the function determined by the relevance determiner 12.

[0067] In the time-series data processing device 1 according to the first embodiment, the data conversion process is a moving average, a difference series, a logarithmic transformation, a seasonal adjustment, or a combination of these. By performing these processes or a combination of these processes on the time-series data, the time-series data processing device 1 can calculate a more accurate extraction relevance degree.

[0068] The time-series data processing method according to the first embodiment includes step ST1 in which an acquisition unit 11 acquires metadata including text information on a dependent variable and attribute information on explanatory variable candidates, and time-series data on the dependent variable and the explanatory variable candidates, and step ST2 in which an association degree determination unit 12 determines a function that defines the relationship between the dependent variable and the explanatory variable candidates, using the metadata and the time-series data. By executing the above method, the time-series data processing device 1 can select explanatory variables that have a statistical association with the time-series data of the dependent variable and a semantic relationship with the dependent variable, as explanatory variables to be used for predicting the dependent variable.

[0069] By executing the program according to the first embodiment, the computer functions as an acquisition unit 11 that acquires metadata including text information on the dependent variable and attribute information on the explanatory variable candidates, and time-series data on the dependent variable and the explanatory variable candidates, and an association determination unit 12 that determines a function that defines the relationship between the dependent variable and the explanatory variable candidates using the metadata and the time-series data. The computer that executes the program can select explanatory variables that have a statistical association with the time-series data of the dependent variable and a semantic relationship with the dependent variable, as explanatory variables to be used for predicting the dependent variable.

[0070] Second Embodiment A time-series data processing device according to a second embodiment performs hierarchical clustering of explanatory variable candidates extracted using the relevance for extraction, using the similarity between the explanatory variable candidates.

[0071] (Basic Configuration of a Time Series Data Processing Device) FIG. 6 is a block diagram showing an example configuration of a time series data processing device 1A according to embodiment 2. The time series data processing device 1A selects explanatory variables to be used for predicting a target variable from among a plurality of types of explanatory variables. As shown in FIG. 6 , the time series data processing device 1A includes an acquisition unit 11A, a relevance determination unit 12A, an extraction unit 13A, a similarity determination unit 14, and a grouping unit 15. For example, the time series data processing device 1A is implemented by a computer. A memory included in the computer stores a program constituting an information processing application for implementing the functions of the acquisition unit 11A, the relevance determination unit 12A, the extraction unit 13A, the similarity determination unit 14, and the grouping unit 15. A processor included in the computer executes the information processing application read from the memory, thereby implementing the functions of the acquisition unit 11A, the relevance determination unit 12A, the extraction unit 13A, the similarity determination unit 14, and the grouping unit 15.

[0072] (Acquisition Unit) The acquisition unit 11A acquires time-series data and metadata. For example, it is assumed that the time-series data processing device 1A includes a communication unit (not shown in FIG. 6 ) and that this communication unit enables communication connection with an external device (not shown in FIG. 6 ). In this case, the acquisition unit 11A controls the communication unit to acquire time-series data and metadata from the external device. The external device is, for example, a data server. Examples of communication methods used for communication with the external device include Bluetooth, Wi-Fi, Zigbee, LoRa, and NFC. The communication unit is, for example, a wireless communication module compatible with any of these communication methods.

[0073] (Relevance Determination Unit) The relevance determination unit 12A determines a function that defines the relationship between the objective variable and the explanatory variable candidates using metadata and time-series data. For example, the relevance determination unit 12A determines a function type based on the objective variable and the explanatory variable candidates from among various function types that calculate numerical information about the relationship between the objective variable and the explanatory variable candidates using metadata and time-series data. The function type may be any type that calculates numerical information when metadata and time-series data are substituted. Determining the function also includes selecting a trained model generated by machine learning.

[0074] The relevance determination unit 12A also determines a function for calculating a degree of relevance for extraction, which is a combination of the statistical relevance of the time-series data between the dependent variable and the explanatory variable candidates and the semantic relevance between the dependent variable and the explanatory variable candidates based on the metadata. This enables the time-series data processing device 1A to select explanatory variables that are semantically related to the dependent variable in addition to the statistical relevance with the time-series data of the dependent variable.

[0075] (Extraction Unit) The extraction unit 13A extracts explanatory variables from the explanatory variable candidates based on a comparison between the function value determined by the relevance determination unit 12A and a threshold. For example, the extraction unit 13A extracts explanatory variable candidates corresponding to function values ​​exceeding a threshold from among multiple explanatory variable candidates. This allows the time series data processing device 1A to extract explanatory variables effective for predicting the dependent variable. The threshold is a threshold related to the function value that is preset for the extraction unit 13A and is determined, for example, based on the prediction result of the dependent variable using the extracted explanatory variables.

[0076] (Similarity Determination Unit) The similarity determination unit 14 determines a function for calculating extraction similarities that combine the statistical similarity of the time-series data between the explanatory variables extracted by the extraction unit 13A and the semantic similarity between the explanatory variables based on the metadata of the explanatory variables extracted by the extraction unit 13A. The statistical similarity is an index indicating the statistical closeness of the time-series data between the explanatory variables extracted by the extraction unit 13A. This similarity is a degree of similarity that takes into account the data trends, small fluctuations, or phase shifts of the waveforms formed by the time-series data of the explanatory variable candidates. The semantic similarity is an index indicating the semantic closeness of the attribute information between the explanatory variable candidates. The attribute information of the explanatory variable candidates is related to the specific field or area to which the explanatory variable candidates belong and can also be considered domain knowledge. Therefore, the semantic similarity is also an index of the domain knowledge similarity of the attribute information between the explanatory variable candidates.

[0077] (Grouping Unit) The grouping unit 15 groups the first extracted explanatory variable data extracted by the extraction unit 13 using the extraction similarity calculated from the function determined by the similarity determination unit 14 and outputs explanatory variables of representative data for each group. Hereinafter, the explanatory variables that serve as representative data for a group will be referred to as second extracted explanatory variable data. Note that the grouping of the first extracted explanatory variable data performed by the grouping unit 15 is hierarchical grouping using the extraction similarity. Furthermore, the second extracted explanatory variable data that serves as representative data for a group is, for example, an explanatory variable with the highest average extraction similarity among other explanatory variables belonging to the same group. For example, the second extracted explanatory variable data extracted by the time series data processing device 1A is output to a prediction device, and the prediction device predicts a target variable using the second extracted explanatory variable data.

[0078] Next, a time-series data processing method according to the second embodiment will be described. FIG. 7 is a flowchart showing the time-series data processing method according to the second embodiment, illustrating a series of processes performed by the time-series data processing device 1A. First, the acquisition unit 11A acquires time-series data related to the dependent variable, time-series data related to the explanatory variable candidates, and metadata including keywords related to the dependent variable and attribute information related to the explanatory variable candidates (step ST1A). For example, the time-series data related to the dependent variable and the time-series data related to the explanatory variable candidates are data sets acquired at regular intervals.

[0079] As the relevance determination process, the relevance determination unit 12A determines a function that defines the relationship between the objective variable and the explanatory variable candidates using the metadata and the time-series data (step ST2A). For example, the relevance determination unit 12A determines a function for calculating a degree of relevance for extraction that is a combination of the statistical relevance of the time-series data between the objective variable and the explanatory variable candidates and the semantic relevance between the objective variable and the explanatory variable candidates based on the metadata.

[0080] The relevance determination unit 12A performs missing value processing, standardization processing, and data conversion processing on the time-series data of the dependent variable and the explanatory variable candidates, as in Embodiment 1, and calculates the value of the function using the processed time-series data. For example, the relevance determination unit 12A performs the above processing on the time-series data related to the dependent variable and the time-series data related to the explanatory variable candidates acquired by the acquisition unit 11A.

[0081] The relevance determination unit 12A calculates the extraction relevance as a function value for the determined function using the processed time-series data. The extraction relevance calculated by the relevance determination unit 12A is output to the extraction unit 13A. The extraction unit 13A extracts first extraction explanatory variable data from the explanatory variable candidates based on a comparison between the extraction relevance and a threshold (step ST3A). For example, the extraction unit 13A extracts explanatory variables corresponding to extraction relevance exceeding the threshold from the multiple explanatory variable candidates.

[0082] The similarity determination unit 14 determines a function for calculating a similarity for extraction that is a combination of the statistical similarity of the time-series data between the explanatory variables extracted by the extraction unit 13A and the semantic similarity between the explanatory variables based on the metadata of the explanatory variables extracted by the extraction unit 13A (step ST4A).The similarity determination unit 14 then calculates a similarity for extraction as a function value for the determined function using the attribute information and time-series data related to the explanatory variable candidates acquired by the acquisition unit 11A.The similarity for extraction calculated by the similarity determination unit 14 is output to the grouping unit 15.

[0083] (Specific Example of Similarity Calculation) The function for calculating the extraction similarity is a function having a first part and a second part. The first part is a part for calculating the correlation coefficient, DTW, or Euclidean distance of the time-series data between the explanatory variable candidates, or a value obtained by combining these, as the statistical similarity of the time-series data between the explanatory variable candidates. The second part is a part for calculating the cosine similarity of the vectorized attribute information of the explanatory variable candidates as the semantic similarity of the attribute information between the explanatory variable candidates.

[0084] (Calculation of Statistical Similarity) In the first part, the statistical similarity of the time series data is calculated for each candidate explanatory variable. For example, the first part is composed of a combination of three indices for calculating statistical similarity: correlation coefficient, DTW, and Euclidean distance. In the first part, the score value obtained for each indices is converted so that it ranges from 0 (low) to 1 (high), and the statistical similarity is calculated by taking a weighted average of these scores.

[0085] The time series data (x, y) of the objective variable and the explanatory variable candidates input to the first part is normalized by the similarity determination unit 14 so that the value of each point falls within the range of 0 to 1 according to the above formula (1). The correlation coefficient, which is an index for calculating statistical similarity, is a statistic that indicates the strength of the linear relationship between the time series data of the explanatory variable candidates. The value of the correlation coefficient ranges from -1 to 1, and the closer it is to 1, the stronger the positive correlation is, and the closer it is to -1, the stronger the negative correlation is. The similarity determination unit 14 calculates the correlation coefficient s for the pair of normalized data x', y' according to the above formula (2). ｃｏｒ Calculate.

[0086] DTW is an index for evaluating the similarity of time series data, and is effective even when the time axes are misaligned. DTW calculates the distance by optimally aligning two time series data. The similarity determination unit 14 calculates a DTW score s for a pair of normalized data x', y', taking into account positive and negative values ​​according to the above formula (3). ｄｔｗ In the above formula (3), N is the length of the time series data.

[0087] The Euclidean distance is the linear distance between two data points, and in the case of time-series data, it is calculated by comparing values ​​at the same time point. The similarity determination unit 14 calculates the Euclidean distance score s for the pair x', y' of normalized data, taking into account positive and negative values, according to the above formula (4). Ｌ２ Calculate.

[0088] The similarity determination unit 14 calculates the weighted average value s of the above three indices as the statistical similarity in accordance with the above formula (5). In the above formula (5), i is a subscript indicating the three indices, namely, the correlation coefficient, the DTW, and the Euclidean distance. s ｉ is the score of the index indicated by the subscript i.

[0089] (Calculation of Semantic Similarity) In the second part, semantic similarity is calculated by quantifying the degree of agreement of attribute information between explanatory variable candidates. Furthermore, if the attribute information related to the explanatory variable candidates is hierarchical information, the semantic similarity is calculated for each layer. For example, in the second part, the attribute information between the explanatory variable candidates is vectorized, and the cosine similarity of the attribute information vector is calculated as the semantic similarity. Vectorization (numerical expression) is achieved by using word2vec, BERT, or the like. However, if the cosine similarity is a negative value, the vector is set to 0.

[0090] FIG. 8A is a diagram showing an example of attribute information related to explanatory variable candidates. In FIG. 8A , explanatory variable a, explanatory variable b, explanatory variable c, ..., explanatory variable h, explanatory variable i, ... are explanatory variable candidates. Explanatory variables a, b, and c correspond to objective variables whose keywords are "construction" and "start of construction". Explanatory variables a and b include information belonging to "construction" and "start of construction", etc. In the attribute information of explanatory variables a and b, the major category is "construction" and the medium category is "start of construction". Explanatory variable c includes information belonging to "construction" and "order", etc. In the attribute information of explanatory variable c, the major category is "construction" and the medium category is "order".

[0091] The explanatory variables h and i correspond to the objective variables whose keywords are "air conditioner," "air conditioner," and "outdoor unit." The explanatory variable h includes information belonging to "machinery" corresponding to "air conditioner," "air conditioner," or "outdoor unit," etc., and includes information related to its "production." As shown in FIG. 8A , the attribute information of the explanatory variable h includes the major category of "machinery" and the medium category of "production." Similarly, the explanatory variable i includes information belonging to "machinery" corresponding to "air conditioner," "air conditioner," or "outdoor unit," etc., and includes information related to its "inventory." The attribute information of the explanatory variable i includes the major category of "machinery" and the medium category of "inventory."

[0092] 8B is a diagram showing an example of attribute pairs between explanatory variable candidates. The attribute pairs shown in FIG. 8B are pairs between attribute i of a certain explanatory variable candidate and attribute j of another explanatory variable candidate. "Order" of attribute i is a pair with "Construction start" of attribute j, and "Sales" of attribute i is a pair with "Inventory" of attribute j. For attribute pairs registered in advance by the user, the semantic similarity between the explanatory variable candidates corresponding to this pair is set to "1."

[0093] 8C is a diagram showing an example of weights assigned to attribute information between explanatory variable candidates. As shown in FIG. 8A, the attribute information regarding the explanatory variable candidates is divided into two hierarchical levels, a major classification and a medium classification, and a weight is assigned to each hierarchical level. In the example shown in FIG. 8C, a weight of 0.75 is assigned to the major classification, and a weight of 0.25 is assigned to the medium classification.

[0094] FIG. 8D is a diagram illustrating an example of semantic similarity. In FIG. 8D , explanatory variable a, explanatory variable b, explanatory variable c, ..., explanatory variable h, explanatory variable i, ... are explanatory variable candidates with attribute information of two levels: major classification and intermediate classification. The similarity determination unit 14 calculates two semantic similarities, one for the major classification and the other for the intermediate classification, for one pair of explanatory variables. For the pair of explanatory variable a and explanatory variable b, the major classification is "construction" and the intermediate classification is "construction commencement." Therefore, the semantic similarity for these pairs is 1. For the pair of explanatory variable a and explanatory variable c and the pair of explanatory variable b and explanatory variable c, the intermediate classifications for explanatory variable a and explanatory variable b are different ("construction commencement") and explanatory variable c are "order received," respectively. However, since these are registered by a user, for example, the semantic similarity is corrected to 1. On the other hand, for example, for the pair of explanatory variable a and explanatory variable h, the degree of agreement for both the major classification and the intermediate classification is low, resulting in a semantic similarity of 0.33.

[0095] When attribute information has multiple hierarchical levels and multiple semantic similarities corresponding to the number of hierarchical levels are calculated for one pair of explanatory variables, the similarity determination unit 14 may calculate a weighted average of these semantic similarities as the semantic similarity for this pair of explanatory variables. Furthermore, in the function determined by the similarity determination unit 14, a weighted average of the statistical similarity calculated by the first part and the semantic similarity calculated by the second part is calculated as the similarity for extraction. The similarity for extraction calculated by the similarity determination unit 14 is output to the grouping unit 15.

[0096] 7, the grouping unit 15 groups the first extracted explanatory variable data extracted by the extraction unit 13A using the extraction similarity calculated from the function determined by the similarity determination unit 14, and outputs explanatory variables that are representative data for each group as second extracted explanatory variable data (step ST5A). The grouping unit 15 performs hierarchical clustering of the first extracted explanatory variable data using the extraction similarity between the explanatory variables.

[0097] For example, the grouping unit 15 may perform agglomerative hierarchical clustering. Agglomerative hierarchical clustering is a clustering method for capturing the hierarchical structure of data. This method starts with each data point as a separate group, and then gradually combines the groups until all data points are finally combined into a single group. This process forms a hierarchical structure of groups.

[0098] 9A is a diagram showing a similarity matrix whose elements are the similarities between each pair of explanatory variables a to d. In FIG. 9A, the similarity between explanatory variable a and explanatory variable b is 0.95, the similarity between explanatory variable a and explanatory variable c is 0.74, and the similarity between explanatory variable a and explanatory variable d is 0.80. The similarity between explanatory variable b and explanatory variable c is 0.75, and the similarity between explanatory variable b and explanatory variable d is 0.96. The similarity between explanatory variable c and explanatory variable d is 0.72.

[0099] The grouping unit 15 may perform clustering using dissimilarity (= 1 - similarity) as an index corresponding to the distance between explanatory variables. FIG. 9B is a diagram showing the distance between groups, in which explanatory variables are clustered according to the similarity shown in FIG. 9A. This diagram shows the clustering results when the shortest distance method is used. The grouping unit 15 can perform clustering of explanatory variables a to d as shown in FIG. 9B by using dissimilarity (= 1 - similarity) as the distance and a threshold value THB (= 0.10) for this distance.

[0100] Based on the clustering results, the grouping unit 15 determines the groups surrounded by dashed lines in FIG. 9B using a similarity threshold THA (=0.90) specified by the user. Since the distances between explanatory variables a, b, and d are less than or equal to the threshold THB, they are grouped into the same group (1). On the other hand, the distance between explanatory variable c (=0.25) is greater than or equal to the threshold THB, so it is grouped into a different group (2) from group (1). The distance between groups can be calculated using the shortest distance method (simple linkage method), the longest distance method (complete linkage method), the group average method, or the like.

[0101] The grouping unit 15 extracts, as second extracted explanatory variable data, an explanatory variable having the highest average extraction similarity compared to other explanatory variables in the group.

[0102] The second extracted explanatory variable data extracted by the time series data processing device 1A is output to a prediction device. The prediction device predicts a dependent variable using the second extracted explanatory variable data. By executing the processes of steps ST1A to ST5A, the time series data processing device 1A can select, as second extracted explanatory variable data to be used in predicting the dependent variable, explanatory variables that are statistically related to the time series data of the dependent variable and that are semantically related to the dependent variable.

[0103] As described above, the time series data processing device 1A according to the second embodiment includes a similarity determination unit 14 that determines a function for calculating a similarity for extraction that is a combination of the statistical similarity of the time series data between the explanatory variables extracted by the extraction unit 13A and the semantic similarity between the explanatory variables based on the metadata of the explanatory variables extracted by the extraction unit 13A, and a grouping unit 15 that groups the explanatory variables extracted by the extraction unit 13 using the similarity for extraction calculated from the function determined by the similarity determination unit 14 and outputs explanatory variables of representative data for each group. This allows the time series data processing device 1A to avoid selecting explanatory variables that are semantically similar in addition to being statistically similar between the explanatory variables.

[0104] In the time series data processing device 1A according to the second embodiment, the function determined by the similarity determiner 14 includes a part for calculating the correlation coefficient, dynamic time warping, or Euclidean distance of the time series data between the explanatory variables, or a combination of these, as the statistical similarity of the time series data between the explanatory variables, and a part for calculating the cosine similarity of the vectorized attribute information of the explanatory variable candidates as the semantic similarity of the attribute information of the explanatory variable candidates. This enables the time series data processing device 1A to avoid selecting semantically similar explanatory variables in addition to the statistical similarity between the explanatory variables.

[0105] Embodiment 3 A time-series data processing device according to embodiment 3 displays information such as waveforms, meta information, relevance, and similar variables of explanatory variables extracted for use in predicting a target variable, thereby presenting this information to a user.

[0106] (Basic Configuration of a Time Series Data Processing Device) FIG. 10 is a block diagram showing an example configuration of a time series data processing device 1B according to embodiment 3. The time series data processing device 1B selects explanatory variables to be used for predicting a target variable from among a plurality of types of explanatory variables. As shown in FIG. 10, the time series data processing device 1B may be a device connected to a display device 2 via a wired or wireless connection. The display device 2 displays information output from the time series data processing device 1B.

[0107] For example, the display device 2 acquires and displays information such as waveforms, meta information, relevance, and similar variables of explanatory variables extracted for use in predicting the dependent variable from the time series data processing device 1B. By referring to the information displayed on the display device 2 and operating an operation device not shown in Fig. 10, the user can select explanatory variables to be used in predicting the dependent variable. The display device 2 may be a display device included in the time series data processing device 1B, or may be a display unit included in a terminal communicatively connected to the time series data processing device 1B via a network.

[0108] 10 , the time series data processing device 1B includes an acquisition unit 11B, a relevance determination unit 12B, an extraction unit 13A, a similarity determination unit 14A, a grouping unit 15A, and an output unit 16. For example, the time series data processing device 1B is realized by a computer. A memory included in the computer stores a program constituting an information processing application for realizing each function of the acquisition unit 11B, the relevance determination unit 12B, the extraction unit 13A, the similarity determination unit 14A, the grouping unit 15A, and the output unit 16. A processor included in the computer executes the information processing application read from the memory, thereby realizing each function of the acquisition unit 11B, the relevance determination unit 12B, the extraction unit 13A, the similarity determination unit 14A, the grouping unit 15A, and the output unit 16.

[0109] (Acquisition Unit) The acquisition unit 11B acquires time-series data and metadata. For example, it is assumed that the time-series data processing device 1B includes a communication unit (not shown in FIG. 10 ) and that this communication unit enables communication with an external device (not shown in FIG. 10 ). In this case, the acquisition unit 11B controls the communication unit to acquire time-series data and metadata from the external device. The external device is, for example, a data server. Examples of communication methods used for communication with the external device include Bluetooth, Wi-Fi, Zigbee, LoRa, and NFC. The communication unit is, for example, a wireless communication module compatible with any of these communication methods. In the third embodiment, the acquisition unit 11B outputs the acquired metadata to the output unit 16 in addition to the relevance determination unit 12B and the similarity determination unit 14A.

[0110] (Relevance Determination Unit) The relevance determination unit 12B determines a function that defines the relationship between the objective variable and the explanatory variable candidates using metadata and time-series data. For example, the relevance determination unit 12B determines a function type based on the objective variable and the explanatory variable candidates from among various function types that calculate numerical information about the relationship between the objective variable and the explanatory variable candidates using metadata and time-series data. The function type may be any type that calculates numerical information when metadata and time-series data are substituted. The determination of the function also includes selecting a trained model generated by machine learning.

[0111] The relevance determiner 12B also determines a function for calculating a degree of relevance for extraction, which is a combination of the statistical relevance of the time-series data between the dependent variable and the candidate explanatory variables and the semantic relevance between the dependent variable and the candidate explanatory variables based on the metadata. Furthermore, the relevance determiner 12B performs missing value processing, standardization processing, and data conversion processing on the time-series data between the dependent variable and the candidate explanatory variables, as in the first embodiment, and calculates a function value using the processed time-series data. The relevance determiner 12B calculates a degree of relevance for extraction as a function value for the determined function using the processed time-series data, and outputs the calculated degree of relevance for extraction to the extractor 13A. In the third embodiment, the relevance determiner 12B outputs the determined function value (degree of relevance for extraction) to the extractor 13A, and also outputs the processed time-series data to the output unit 16.

[0112] (Extraction Unit) The extraction unit 13A extracts explanatory variables from the explanatory variable candidates based on a comparison between the function value determined by the relevance determination unit 12B and a threshold. For example, the extraction unit 13A extracts explanatory variable candidates corresponding to function values ​​exceeding a threshold from among multiple explanatory variable candidates. This allows the time series data processing device 1B to extract explanatory variables effective for predicting the dependent variable. The threshold is a threshold related to the function value that is preset for the extraction unit 13A and is determined, for example, based on the prediction result of the dependent variable using the extracted explanatory variables.

[0113] (Similarity Determination Unit) The similarity determination unit 14A determines a function for calculating extraction similarity, which is a combination of the statistical similarity of the time-series data between the explanatory variables extracted by the extraction unit 13A and the semantic similarity between the explanatory variables based on the metadata of the explanatory variables extracted by the extraction unit 13A. The statistical similarity is an index indicating the statistical closeness of the time-series data between the explanatory variables extracted by the extraction unit 13A. This similarity is a degree of similarity that takes into account the data trend, small fluctuations, or phase shifts of the waveforms formed by the time-series data of the explanatory variable candidates. The semantic similarity is an index indicating the semantic closeness of the attribute information between the explanatory variable candidates. The attribute information related to the explanatory variable candidates is related to the specific field or area to which the explanatory variable candidate belongs and can also be considered domain knowledge. Therefore, the semantic similarity is also an index of the domain knowledge similarity of the attribute information between the explanatory variable candidates. In the third embodiment, the similarity determination unit 14A adds the determined function value (extraction similarity) to the grouping unit 15A and also outputs it to the output unit 16.

[0114] (Grouping Unit) The grouping unit 15A groups the first extracted explanatory variable data extracted by the extraction unit 13A using the extraction similarity calculated from the function determined by the similarity determination unit 14A, and outputs explanatory variables of representative data for each group. Hereinafter, the explanatory variables that serve as representative data for a group will be referred to as second extracted explanatory variable data. Note that the grouping of the first extracted explanatory variable data performed by the grouping unit 15A is hierarchical grouping using the extraction similarity. Furthermore, the second extracted explanatory variable data that serves as representative data for a group is, for example, an explanatory variable with the highest average extraction similarity among other explanatory variables belonging to the same group. In the third embodiment, the second extracted explanatory variable data extracted by the grouping unit 15A is output to the output unit 16.

[0115] (Output Unit) The output unit 16 is a so-called visualization unit that displays the waveforms, metadata, relevance, or similarity related to the explanatory variables extracted by the extraction unit 13A on the display device 2. For example, the output unit 16 generates display information for displaying the second extracted explanatory variable data, the waveforms of the time-series data, the metadata, the relevance for extraction, or the similarity for extraction extracted by the grouping unit 15A, and outputs the display information to the display device 2. The display device 2 displays the second extracted explanatory variable data, the waveforms of the time-series data, the metadata, the relevance for extraction, or the similarity for extraction based on the display information acquired from the output unit 16.

[0116] (Specific examples of display screens) The output unit 16 causes the display device 2 to display, for example, an explanatory variable list display screen including the second extracted explanatory variable data, the waveform of the time-series data, the metadata, the relevance for extraction or the similarity for extraction, an individual explanatory variable display screen, or a selected explanatory variable display screen.

[0117] FIG. 11 is a screen diagram showing a screen example (1) in embodiment 3, illustrating an explanatory variable list display screen 2A. The output unit 16 generates display information for displaying the explanatory variable list display screen 2A and outputs it to the display device 2. The display device 2 displays the explanatory variable list display screen 2A based on the display information acquired from the output unit 16. Various information related to the explanatory variables can be selected from the explanatory variable list display screen 2A by operating an operation device (not shown in FIG. 10). The operation device may be included in the time series data processing device 1B, or may be an operation unit included in a terminal communicatively connected to the time series data processing device 1B via a network.

[0118] As shown in FIG. 11 , the explanatory variable list display screen 2A displays setting fields for displaying information set by the user. These setting fields display, for example, a “Number of Selected Explanatory Variables” field 21, a “Number of Displayed Similar Variables” field 22, and a “Similarity Threshold” field 23. The “Number of Selected Explanatory Variables” field 21 sets the number of explanatory variables currently selected on the explanatory variable list display screen 2A (e.g., 10). The “Number of Displayed Similar Variables” field 22 sets the maximum number of similar explanatory variables that can be displayed on the explanatory variable list display screen 2A (e.g., 3). The “Similarity Threshold” field 23 sets a similarity threshold (e.g., 0.75) for determining which explanatory variables are to be displayed on the explanatory variable list display screen 2A. The explanatory variable list display screen 2A displays explanatory variables (second extracted explanatory variable data) whose extraction similarity is equal to or greater than the similarity threshold, in descending order of extraction similarity, up to the number set in the “Number of Displayed Similar Variables” field 22.

[0119] As shown in FIG. 11 , the explanatory variable list display screen 2A displays a "Selection" column, a "Meta Information" column, a "Data Name" column, a "Relevance" column, and a "Similar Variable" column related to the second extracted explanatory variable data extracted by the grouping unit 15. An explanatory variable is selected by checking a check box in the "Selection" column using the operation device. On the explanatory variable list display screen 2A shown in FIG. 11 , the check box 24 is checked, so explanatory variable a is selected, and the check box 25 is checked, so explanatory variable b is selected. Furthermore, because explanatory variable b is selected, the check box 26 is also checked accordingly.

[0120] The attribute information, "data name," and "relevance" displayed in the "major classification" and "middle classification" of the "meta information" column 27 can be selected using the operating device to sort the data in text order (e.g., A-Z) if the column contents are text, or in descending order (e.g., relevance) if the column contents are numerical values. Also, by checking a checkbox 28 using the operating device, explanatory variables similar to explanatory variable a are displayed in the display column 29 in the number set in the "number of similar variables to display" column 22. For example, information on explanatory variables b, g, and m that are not classified in the same group as explanatory variable a but are similar in terms of statistical or metadata semantics of the time-series data is displayed in the display column 29.

[0121] Furthermore, when a cell containing "explanatory variable a" is selected using the operating device, the operation information is output to the output unit 16. In response to the operation information, the output unit 16 generates display information for displaying an explanatory variable individual display screen 2B related to explanatory variable a, and outputs the generated display information to the display device 2. Based on this display information, the display device 2 displays the explanatory variable individual display screen 2B related to explanatory variable a.

[0122] FIG. 12 is a screen diagram showing a screen example (2) in embodiment 3, illustrating an explanatory variable individual display screen 2B. The output unit 16 generates display information for displaying an explanatory variable individual display screen 2B for one explanatory variable, and outputs it to the display device 2. The display device 2 displays the explanatory variable individual display screen 2B based on the display information acquired from the output unit 16. The explanatory variable individual display screen 2B displays various information related to an explanatory variable selected using the operation device. For example, it is assumed that the cell for "explanatory variable a" is selected on the explanatory variable list display screen 2A, thereby displaying the explanatory variable individual display screen 2B related to explanatory variable a.

[0123] As shown in FIG. 12 , the explanatory variable individual display screen 2B displays a setting field 21A displaying information set by the user, an explanatory variable information field 22A, and a waveform field 23A. The setting field 21A displays a “Number of Selected Explanatory Variables,” a “Number of Displayed Similar Variables,” a “Similarity Threshold,” and a “Similar Variable Waveform Display” field with a check box 211A. The “Number of Selected Explanatory Variables” field sets the number of explanatory variables currently selected in the explanatory variable individual display screen 2B (e.g., 10). The “Number of Displayed Similar Variables” field sets the maximum number of similar explanatory variables that can be displayed in the explanatory variable individual display screen 2B (e.g., 3). The “Similarity Threshold” field sets a similarity threshold (e.g., 0.75) for determining which explanatory variables to display in the explanatory variable individual display screen 2B. For each explanatory variable, the explanatory variable individual display screen 2B displays explanatory variables (second extracted explanatory variable data) whose extraction similarity is equal to or greater than the similarity threshold, in descending order of extraction similarity, up to the number set in the “Number of Displayed Similar Variables” field.

[0124] The waveform 231A of the time-series data of explanatory variable a selected for individual display and the corresponding objective variable are always displayed. Cell 221A in the explanatory variable information field 22A displays "explanatory variable a," which is the explanatory variable to be individually displayed. Checking checkbox 211A displays the waveforms of similar explanatory variables b, g, and m linked to explanatory variable a. Since explanatory variable b is selected on the explanatory variable list display screen 2A shown in FIG. 11 , the checkbox 222A in the selection field in the explanatory variable information field 22A is also checked in conjunction with the selection. Display field 223A displays information on explanatory variables b, g, and m that are not classified into the same group as explanatory variable a but are semantically similar in terms of statistical or metadata of the time-series data.

[0125] FIG. 13 is a screen diagram showing a screen example (3) in the third embodiment, illustrating a selected explanatory variable display screen 2C. The output unit 16 generates display information for displaying the selected explanatory variable display screen 2C and outputs it to the display device 2. The display device 2 displays the selected explanatory variable display screen 2C based on the display information acquired from the output unit 16. The selected explanatory variable display screen 2C displays various information related to explanatory variables selected using the operation device. For example, by checking multiple check boxes in the selection column on the explanatory variable list display screen 2A, the selected explanatory variable display screen 2C related to the explanatory variables selected by these check boxes is displayed.

[0126] As shown in FIG. 13 , the selected explanatory variable display screen 2C displays a setting field 21B, which displays information set by the user, an explanatory variable information field 22B, and a waveform field 23B. The setting field 21B sets the number of selected explanatory variables. The selected explanatory variable display screen 2C displays the waveform, relevance, meta information, and data name for the selected explanatory variable. In the initial display, all check boxes in the selection fields of the explanatory variable information field 22B are checked. The waveform field 23B displays the waveforms of the target variable and the time-series data of the explanatory variables whose selection fields in the explanatory variable information field 22B are checked. Furthermore, explanatory variables can be selected from the selected explanatory variable display screen 2C. For example, explanatory variables can be prevented from being used by unchecking the check box in the “Select” field of the explanatory variable information field 22B corresponding to the explanatory variable that is determined not to be used. In this case, the waveforms of the time-series data corresponding to the unchecked explanatory variables are hidden in the waveform field 23B. As described above, by displaying the explanatory variable list display screen 2A, the individual explanatory variable display screen 2B, or the selected explanatory variable display screen 2C on the display device 2, a user who refers to these screens can efficiently select explanatory variables to be used in predicting the target variable.

[0127] As described above, the time-series data processing device 1B according to the third embodiment includes the output unit 16 that displays the waveform, metadata, relevance, or similarity related to the explanatory variables extracted by the extraction unit 13A on the display device 2. By referring to the information displayed on the display device 2, the user can efficiently select explanatory variables to be used for predicting the target variable.

[0128] Fourth Embodiment A time-series data processing device according to a fourth embodiment extracts explanatory variables associated with a target variable by an independence test, even when other explanatory variables are conditioned.

[0129] (Basic Configuration of a Time Series Data Processing Device) FIG. 14 is a block diagram showing an example configuration of a time series data processing device 1C according to embodiment 4. The time series data processing device 1C selects explanatory variables to be used for predicting a target variable from among a plurality of types of explanatory variables. As shown in FIG. 14, the time series data processing device 1C may be a device connected to a display device 2 via a wired or wireless connection. The display device 2 displays information output from the time series data processing device 1C.

[0130] For example, the display device 2 acquires and displays information such as waveforms, meta information, relevance, and similar variables of explanatory variables extracted for use in predicting the dependent variable from the time series data processing device 1C. By referring to the information displayed on the display device 2 and operating an operation device not shown in Fig. 14, the user can select explanatory variables to be used in predicting the dependent variable. Note that the display device 2 may be a display device included in the time series data processing device 1C, or may be a display unit included in a terminal communicatively connected to the time series data processing device 1C via a network.

[0131] The time series data processing device 1C has a configuration in which the testing unit 17 is added to the time series data processing device 1B, but is not limited to this. For example, the time series data processing device 1C may have a configuration in which the testing unit 17 is added to the time series data processing device 1 or the time series data processing device 1A.

[0132] The time series data processing device 1C includes an acquisition unit 11B, a relevance determination unit 12C, an extraction unit 13A, a similarity determination unit 14B, a grouping unit 15B, an output unit 16A, and a testing unit 17. For example, the time series data processing device 1C is realized by a computer. A memory included in the computer stores a program constituting an information processing application for realizing each of the functions of the acquisition unit 11B, the relevance determination unit 12C, the extraction unit 13A, the similarity determination unit 14B, the grouping unit 15B, the output unit 16A, and the testing unit 17. A processor included in the computer executes the information processing application read from the memory, thereby realizing each of the functions of the acquisition unit 11B, the relevance determination unit 12C, the extraction unit 13A, the similarity determination unit 14B, the grouping unit 15B, the output unit 16A, and the testing unit 17.

[0133] (Acquisition Unit) The acquisition unit 11B acquires time-series data and metadata. For example, assume that the time-series data processing device 1C includes a communication unit (not shown in FIG. 14 ) and is capable of communicating with an external device (not shown in FIG. 14 ) via this communication unit. In this case, the acquisition unit 11B controls the communication unit to acquire time-series data and metadata from the external device. The external device is, for example, a data server. Examples of communication methods used for communication with the external device include Bluetooth, Wi-Fi, Zigbee, LoRa, and NFC. The communication unit is, for example, a wireless communication module compatible with any of these communication methods. In the fourth embodiment, the acquisition unit 11B outputs the acquired metadata to the relevance determination unit 12C, the similarity determination unit 14B, and the output unit 16A, as well as to the testing unit 17.

[0134] (Relevance Determination Unit) The relevance determination unit 12C determines a function that defines the relationship between the objective variable and the explanatory variable candidates using metadata and time-series data. For example, the relevance determination unit 12C determines a function type based on the objective variable and the explanatory variable candidates from among various function types that calculate numerical information about the relationship between the objective variable and the explanatory variable candidates using metadata and time-series data. The function type may be any type that calculates numerical information when metadata and time-series data are substituted. Determining the function also includes selecting a trained model generated by machine learning.

[0135] The relevance determiner 12C also determines a function for calculating a degree of relevance for extraction, which is a combination of the statistical relevance of the time-series data between the dependent variable and the candidate explanatory variables and the semantic relevance between the dependent variable and the candidate explanatory variables based on the metadata. Furthermore, the relevance determiner 12C performs missing value processing, standardization processing, and data conversion processing on the time-series data between the dependent variable and the candidate explanatory variables, as in the first embodiment, and calculates a function value using the processed time-series data. The relevance determiner 12C calculates a degree of relevance for extraction as a function value for the determined function using the processed time-series data, and outputs the calculated degree of relevance for extraction to the extractor 13A. In the fourth embodiment, the relevance determiner 12C not only outputs the determined function value (degree of relevance for extraction) to the extractor 13A, but also outputs the processed time-series data to the output unit 16A and the testing unit 17.

[0136] (Extraction Unit) The extraction unit 13A extracts explanatory variables from the explanatory variable candidates based on a comparison between the function value determined by the relevance determination unit 12C and a threshold. For example, the extraction unit 13A extracts explanatory variable candidates corresponding to function values ​​exceeding a threshold from among multiple explanatory variable candidates. This allows the time series data processing device 1C to extract explanatory variables effective for predicting the dependent variable. The threshold is a threshold related to the function value that is preset for the extraction unit 13A and is determined, for example, based on the prediction result of the dependent variable using the extracted explanatory variables.

[0137] (Similarity Determination Unit) The similarity determination unit 14B determines a function for calculating extraction similarity, which is a combination of the statistical similarity of the time-series data between the explanatory variables extracted by the extraction unit 13A and the semantic similarity between the explanatory variables based on the metadata of the explanatory variables extracted by the extraction unit 13A. The statistical similarity is an index indicating the statistical closeness of the time-series data between the explanatory variables extracted by the extraction unit 13A. This similarity is a degree of similarity that takes into account the data trend, small fluctuations, or phase shifts of the waveforms formed by the time-series data of the explanatory variable candidates. The semantic similarity is an index indicating the semantic closeness of the attribute information between the explanatory variable candidates. The attribute information related to the explanatory variable candidates is related to the specific field or area to which the explanatory variable candidate belongs and can also be considered domain knowledge. Therefore, the semantic similarity is also an index of the domain knowledge similarity of the attribute information between the explanatory variable candidates. In the fourth embodiment, the similarity determination unit 14B adds the determined function value (extraction similarity) to the grouping unit 15B and also outputs it to the output unit 16A.

[0138] (Grouping Unit) The grouping unit 15B groups the first extracted explanatory variable data extracted by the extraction unit 13A using the similarity for extraction calculated from the function determined by the similarity determination unit 14B, and outputs explanatory variables of representative data for each group. Hereinafter, the explanatory variables that serve as representative data for a group will be referred to as second extracted explanatory variable data. Note that the grouping of the first extracted explanatory variable data performed by the grouping unit 15B is hierarchical grouping using the similarity for extraction.

[0139] Furthermore, the second extracted explanatory variable data that serves as representative data for the group is, for example, an explanatory variable that has the highest average extraction similarity among other explanatory variables that belong to the same group. In the fourth embodiment, the second extracted explanatory variable data extracted by the grouping unit 15B is output to the testing unit 17.

[0140] (Output Unit) The output unit 16A is a so-called visualization unit that displays, on the display device 2, waveforms, metadata, relevance, or similarity related to the explanatory variables extracted by the testing unit 17 (hereinafter referred to as third extracted explanatory variable data). For example, the output unit 16A generates display information for displaying the third extracted explanatory variable data, waveforms of the time-series data, metadata, relevance for extraction, or similarity for extraction extracted by the testing unit 17, and outputs the display information to the display device 2. The display device 2 displays the third extracted explanatory variable data, waveforms of the time-series data, metadata, relevance for extraction, or similarity for extraction based on the display information acquired from the output unit 16.

[0141] (Testing Unit) The testing unit 17 extracts explanatory variables associated with the target variable through an independence test. For example, the testing unit 17 evaluates the extraction relevance of the second extracted explanatory variable data extracted by the grouping unit 15B using a conditional independence test when other explanatory variables are conditioned. The conditional independence test is a statistical method for evaluating whether a certain explanatory variable depends on other variables. In particular, the conditional independence test is used to evaluate the dependency when other explanatory variables (covariates) are conditioned. The conditional independence test evaluates whether specific variables X and Y are independent, with other variable (covariate) Z as a condition.

[0142] The testing unit 17 uses a conditional independence test on the second extracted explanatory variable data to calculate a p-value that indicates the relevance when other explanatory variables are conditioned according to the following formula (6), and extracts explanatory variables whose p-value is equal to or greater than a threshold as third extracted explanatory variable data. In the following formula (6), S is a set of indexes of explanatory variables after grouping, y is a response variable, and x ｉ is an explanatory variable. ci is a conditional independence test value, which is a function based on, for example, partial correlation coefficient, conditional mutual information, etc. ｉ is the response variable y and the explanatory variable x ｉ The closer to 0, the less independent the test statistic is. ｔｈ is the threshold value of the p-value. ｉ = ci(y, x ｉ , {xｊ |j∈S\i}) (6)

[0143] Next, a description will be given of the operation of the testing unit 17. Fig. 15 is a flowchart showing the operation of the testing unit 17. The testing unit 17 calculates a test statistic according to the above formula (6) using the time-series data processed by the relevance determination unit 12C and the second extracted explanatory variable data extracted by the grouping unit 15B (step ST1B).

[0144] Next, the testing unit 17 calculates the objective variable y and the explanatory variable x ｉ It is checked whether the p-value of the test statistic has been calculated for all pairs with (step ST2B). If there is a pair for which the test statistic has not been calculated (step ST2B; NO), the process returns to step ST1B and the test statistic for the remaining pairs is calculated. On the other hand, if the test statistic has been calculated for all pairs (step ST2B; YES), the p-value is calculated according to the above formula (6). ｉ has been calculated.

[0145] The testing unit 17 ｉ Maximum value of < threshold p ｔｈ It is checked whether or not the value is p (step ST3B). ｉ Maximum value of threshold p ｔｈ If p is ≠ ... ｉ The explanatory variable i corresponding to the maximum value of p is removed (step ST4B), and the process returns to step ST1B. ｉ Maximum value of < threshold p ｔｈ If so (step ST3B; YES), the testing unit 17 ends the series of processes shown in FIG. 15 and sets the explanatory variables corresponding to the indexes included in the index set S as third extracted explanatory variable data.

[0146] As described above, the time series data processing device 1C according to the fourth embodiment includes the testing unit 17 that extracts explanatory variables associated with a target variable by an independence test. By extracting explanatory variables by an independence test, the time series data processing device 1C can select statistically valid explanatory variables.

[0147] Various aspects of the present disclosure are summarized below as appendices.

[0148] (Supplementary Note 1) A time-series data processing device comprising: an acquisition unit that acquires metadata including text information on a dependent variable and attribute information on explanatory variable candidates, and time-series data of the dependent variable and the explanatory variable candidates; and an association determination unit that determines a function defining a relationship between the dependent variable and the explanatory variable candidates using the metadata and the time-series data. (Supplementary Note 2) The time-series data processing device according to Supplementary Note 1, further comprising: an extraction unit that extracts explanatory variables from the explanatory variable candidates based on a comparison between a value of the function determined by the association determination unit and a threshold. (Supplementary Note 3) The time-series data processing device according to Supplementary Note 2, further comprising: an association determination unit that determines a function for calculating an association for extraction formed by a combination of a statistical association of the time-series data between the dependent variable and the explanatory variable candidates and a semantic association between the dependent variable and the explanatory variable candidates based on the metadata; and the extraction unit that extracts the explanatory variables from the explanatory variable candidates based on a comparison between the association for extraction calculated from the function determined by the association determination unit and the threshold. (Supplementary Note 4) The time series data processing device according to Supplementary Note 3, comprising: a similarity determination unit that determines a function for calculating a similarity for extraction, the similarity being a combination of a statistical similarity of the time series data between the explanatory variables extracted by the extraction unit and a semantic similarity between the explanatory variables based on the metadata of the explanatory variables extracted by the extraction unit; and a grouping unit that groups the explanatory variables extracted by the extraction unit using the similarity for extraction calculated from the function determined by the similarity determination unit, and outputs the explanatory variables of representative data for each group. (Supplementary Note 5) The time series data processing device according to Supplementary Note 3 or Supplementary Note 4, comprising: an output unit that displays, on a display device, a waveform, the metadata, the degree of association, or the similarity related to the explanatory variables extracted by the extraction unit. (Supplementary Note 6) The time series data processing device according to any one of Supplementary Notes 1 to 5, comprising: a testing unit that extracts the explanatory variables associated with the target variable by an independence test.(Supplementary Note 7) The time-series data processing device according to any one of Supplementary Notes 1 to 3, wherein the function determined by the relevance determiner comprises: a portion for calculating a correlation coefficient, dynamic time warping, or Euclidean distance of the time-series data between the dependent variable and the explanatory variable candidate, or a value obtained by combining these, as a statistical relevance of the time-series data between the dependent variable and the explanatory variable candidate; and a portion for calculating a cosine similarity between vectorized text information of the dependent variable and attribute information of the explanatory variable candidate, as a semantic relevance between the text information of the dependent variable and the attribute information of the explanatory variable candidate. (Supplementary Note 8) The time-series data processing device according to Supplementary Note 4, wherein the function determined by the similarity determiner comprises: a portion for calculating a correlation coefficient, dynamic time warping, or Euclidean distance of the time-series data between the explanatory variables, or a value obtained by combining these, as a statistical similarity of the time-series data between the explanatory variables; and a portion for calculating a cosine similarity of vectorized attribute information of the explanatory variable candidate, as a semantic similarity of the attribute information of the explanatory variable candidate. (Supplementary Note 9) The time series data processing device according to any one of Supplementary Notes 1 to 8, wherein the relevance determination unit performs missing value processing, standardization processing, and data conversion processing on the time series data of the dependent variable and the explanatory variable candidate acquired by the acquisition unit, and calculates a function value using the processed time series data. (Supplementary Note 10) The time series data processing device according to Supplementary Note 9, wherein the data conversion processing is difference series, logarithmic transformation, seasonal adjustment, or a combination thereof. (Supplementary Note 11) A time series data processing method by a time series data processing device, comprising: an acquisition unit acquiring metadata including text information on the dependent variable and attribute information on the explanatory variable candidate, and time series data of the dependent variable and the explanatory variable candidate; and an relevance determination unit determining a function defining a relationship between the dependent variable and the explanatory variable candidate using the metadata and the time series data.(Supplementary Note 12) A program for causing a computer to function as: an acquisition unit that acquires metadata including text information on a dependent variable and attribute information on explanatory variable candidates, and time-series data on the dependent variable and the explanatory variable candidates; and a relevance determination unit that determines a function that defines a relationship between the dependent variable and the explanatory variable candidates, using the metadata and the time-series data.

[0149] It is possible to combine the embodiments, modify any of the components of the embodiments, or omit any of the components of the embodiments.

[0150] The time-series data processing device according to the present disclosure can be used in, for example, various systems that handle time-series data.

[0151] 1, 1A, 1B, 1C Time series data processing device, 2 Display device, 2A Explanatory variable list display screen, 2B Explanatory variable individual display screen, 2C Selected explanatory variable display screen, 11, 11A, 11B Acquisition unit, 12, 12A, 12B, 12C Association determination unit, 13, 13A Extraction unit, 14, 14A, 14B Similarity determination unit, 15, 15A, 15B Grouping unit, 16, 16A Output unit, 17 Test unit, 21, 21A, 21B, 22, 22A, 22B, 23, 23A, 23B, 27, 29, 223A Column, 24, 25, 26, 28, 211A, 222A Check box, 100 Input interface, 101 Output interface, 102 Processing circuit, 103 Processor, 104 Memory, 221A Cell, 231A waveform.

Claims

1. A time-series data processing device comprising: an acquisition unit that acquires metadata including text information on a dependent variable and attribute information on explanatory variable candidates, and time-series data on the dependent variable and the explanatory variable candidates; and a relevance determination unit that determines a function that defines the relationship between the dependent variable and the explanatory variable candidates using the metadata and the time-series data.

2. The time series data processing device according to claim 1, further comprising an extraction unit that extracts explanatory variables from the explanatory variable candidates based on a comparison between the value of the function determined by the relevance determination unit and a threshold value.

3. The time series data processing device according to claim 2, characterized in that the relevance determination unit determines a function for calculating a relevance for extraction consisting of a combination of a statistical relevance of the time series data between the dependent variable and the candidate explanatory variable and a semantic relevance between the dependent variable and the candidate explanatory variable based on the metadata, and the extraction unit extracts the explanatory variable from the candidate explanatory variable based on a comparison between the relevance for extraction calculated from the function determined by the relevance determination unit and the threshold value.

4. The time series data processing device according to claim 3, further comprising: a similarity determination unit that determines a function for calculating a similarity for extraction consisting of a combination of the statistical similarity of the time series data between the explanatory variables extracted by the extraction unit and the semantic similarity between the explanatory variables based on the metadata of the explanatory variables extracted by the extraction unit; and a grouping unit that uses the similarity for extraction calculated from the function determined by the similarity determination unit to group the explanatory variables extracted by the extraction unit and output the explanatory variables of representative data for each group.

5. A time series data processing device as described in claim 4, characterized in that it is provided with an output unit that displays on a display device the waveforms, the metadata, the relevance or the similarity related to the explanatory variables extracted by the extraction unit.

6. The time series data processing device according to claim 5, further comprising a testing unit that extracts the explanatory variables associated with the target variable by an independence test.

7. The time series data processing device according to any one of claims 1 to 3, characterized in that the function determined by the relevance determination unit has: a part for calculating a correlation coefficient, dynamic time warping or Euclidean distance of the time series data between the dependent variable and the candidate explanatory variable, or a value combining these, as a statistical relevance of the time series data between the dependent variable and the candidate explanatory variable; and a part for calculating a cosine similarity between vectorized text information of the dependent variable and attribute information of the candidate explanatory variable as a semantic relevance between the text information of the dependent variable and the attribute information of the candidate explanatory variable.

8. The time series data processing device according to claim 4, characterized in that the function determined by the similarity determination unit has: a part for calculating a correlation coefficient of the time series data between the explanatory variables, dynamic time warping or Euclidean distance, or a value combining these, as the statistical similarity of the time series data between the explanatory variables; and a part for calculating a cosine similarity of the vectorized attribute information of the explanatory variable candidates as the semantic similarity of the attribute information of the explanatory variable candidates.

9. A time series data processing device according to any one of claims 1 to 8, characterized in that the relevance determination unit performs missing value processing, standardization processing, and data conversion processing on the time series data of the target variable and the candidate explanatory variables acquired by the acquisition unit, and calculates the value of a function using the processed time series data.

10. The time series data processing device according to claim 9, wherein the data transformation process is differential series, logarithmic transformation, seasonal adjustment, or a combination thereof.

11. A time series data processing method by a time series data processing device, comprising: an acquisition unit acquiring metadata including text information on a dependent variable and attribute information on candidate explanatory variables, and time series data on the dependent variable and the candidate explanatory variables; and a relevance determination unit determining a function that defines the relationship between the dependent variable and the candidate explanatory variables using the metadata and the time series data.

12. A program for causing a computer to function as: an acquisition unit that acquires metadata including text information on a dependent variable and attribute information on candidate explanatory variables, and time-series data on the dependent variable and the candidate explanatory variables; and a relevance determination unit that determines a function that defines the relationship between the dependent variable and the candidate explanatory variables using the metadata and the time-series data.

Images