Time series data processing device and time series data processing method

Abstract

A time-series data processing device according to the present invention comprises: a data input unit (110) that acquires a first time-series data set including a plurality of pieces of time-series data that serve as candidates for an explanatory variable; a pre-processing unit (120) that converts, into a calculable format, the relationships among the plurality of pieces of time-series data included in the acquired first time-series data set, and that generates a second time-series data set including the plurality of pieces of time-series data after conversion; a relationship calculation unit (140) that calculates the waveform semantic relationships among the plurality of pieces of time-series data included in the second time-series data set; a hierarchization unit (150) that determines the temporal before-after relationships among the plurality of pieces of time-series data included in the second time-series data set, hierarchizes the plurality of pieces of time-series data included in the second time-series data set in accordance with the determined temporal before-after relationships, and outputs the hierarchization result; and a visualization unit (160) that generates a visualized diagram in which the plurality of pieces of time-series data included in the second time-series data set are visualized on the basis of the calculated waveform semantic relationships and output hierarchization result.

Description

[Technical Field]

[0001] The present disclosure relates to time series data processing techniques. [Background technology]

[0002] In time-series data processing using a prediction model, the selection of explanatory variables is extremely important and greatly influences the accuracy of the prediction. However, it is difficult to manually select the optimal explanatory variable from among a large number of candidate explanatory variables. For this reason, a technology has been proposed that automatically selects explanatory variables in prediction according to a predetermined algorithm (for example, Patent Document 1). In the technology of Patent Document 1, the larger the regression coefficient (absolute value) between the objective variable and the explanatory variable, the more appropriate the explanatory variable is considered, and the explanatory variable is automatically selected by comparing the regression coefficient with a threshold (Claim 4, paragraph 0093 of Patent Document 1). [Prior art documents] [Patent documents]

[0003] [Patent Document 1] International Publication No. 2013 / 187295 Summary of the Invention [Problem to be solved by the invention]

[0004] The technology disclosed in Patent Document 1 for automatically selecting explanatory variables has a problem in that only explanatory variables that have a spurious correlation with the dependent variable may be selected. That is, an explanatory variable that cannot be said to have a causal relationship with the dependent variable may be determined to have a causal relationship due to some factor, and only such explanatory variables may be selected.

[0005] The present disclosure has been made to solve such problems, and aims to provide a time-series data processing technique that can prevent only explanatory variables with spurious correlations from being selected. [Means for solving the problem]

[0006] One aspect of a time series data processing device according to an embodiment of the present disclosure includes: a data input unit that acquires a first time series dataset including a plurality of time series data to be used as candidates for explanatory variables; a preprocessing unit that converts relationships between the plurality of time series data included in the acquired first time series dataset into a calculable format and generates a second time series dataset including the converted plurality of time series data; a relationship calculation unit that calculates waveform semantic relationships between the plurality of time series data included in the second time series dataset; a hierarchizing unit that determines temporal relationships between the plurality of time series data included in the second time series dataset, hierarchizes the plurality of time series data included in the second time series dataset according to the determined temporal relationships, and outputs the hierarchizing results; and a visualization unit that generates a visualization diagram that visualizes the plurality of time series data included in the second time series dataset based on the calculated waveform semantic relationships and the output hierarchizing results. [Effects of the Invention]

[0007] According to the time-series data processing device according to the embodiment of the present disclosure, multiple explanatory variable candidates are presented, so that it is possible to prevent only explanatory variables with spurious correlations from being selected. [Brief explanation of the drawings]

[0008] [Figure 1] 1 is a diagram illustrating an example of the configuration of a time-series data processing device and a time-series data processing system. [Figure 2] FIG. 10 is a diagram illustrating an outline of processing by a grouping unit. [Figure 3] FIG. 10 is a diagram illustrating an overview of relationship calculation performed by a relationship calculation unit. [Figure 4] FIG. 2 is a diagram illustrating an overview of layering performed by a layering unit. [Figure 5] 5A and 5B are diagrams showing variations in layering. Specifically, they are diagrams showing an example of month-shifted correlation. Fig. 5A shows the original waveform. Fig. 5B shows the waveform when one of the waveforms is shifted by one month. [Figure 6] 10A and 10B are diagrams showing variations in layering, specifically, diagrams showing specific examples of layering in accordance with a pre-set temporal context; [Figure 7A] FIG. 1 is a diagram illustrating an example of the hardware configuration of a time-series data processing device. [Figure 7B] FIG. 1 is a diagram illustrating an example of the hardware configuration of a time-series data processing device. [Figure 8] 1 is a flowchart of a time-series data processing method. [Figure 9] FIG. 10 is a diagram illustrating an example of a visualization diagram. [Figure 10] FIG. 10 is a diagram illustrating an outline of a process performed by a recalculation unit. DETAILED DESCRIPTION OF THE INVENTION

[0009] Various embodiments of the present disclosure will be described in detail below with reference to the accompanying drawings. Components with the same or similar reference numerals in the drawings have the same or similar configurations or functions, and redundant descriptions of such components will be omitted. Furthermore, in this disclosure, the term "or" is used to mean an inclusive logical OR unless otherwise specified.

[0010] Furthermore, the term "causal relationship" used in this disclosure refers to a temporal relationship selected by a user from among the temporal relationships determined by statistical analysis of two pieces of time-series data. The user selects a temporal relationship that is convincing to the user.

[0011] Embodiment 1 <Configuration> A time series data processing device and a time series data processing system according to a first embodiment of the present disclosure will be described with reference to Fig. 1. The time series data processing system shown in Fig. 1 includes a time series data processing device 100, a storage device 200, and a storage device 300. The storage device 200 is a device that stores time series data that are candidates for explanatory variables. The storage device 300 is a device that stores additional data.

[0012] (Time series data processing device) The time series data processing device 100 a data input unit 110 that acquires a first time series data set from a storage device 200, the first time series data set including a plurality of time series data that are candidates for explanatory variables; and a pre-processing unit 120 that converts the relationships between the plurality of time series data included in the acquired first time series data set into a calculable format and generates a second time series data set including the plurality of converted time series data. The time series computing system includes a grouping unit (130) that classifies the plurality of time series data included in the second time series data set into groups according to the properties of the data and generates a representative value for each group; a relationship calculation unit (140) that calculates a waveform semantic relationship between the plurality of time series data included in the second time series data set; a hierarchizing unit (150) that determines a temporal relationship between the plurality of time series data included in the second time series data set, hierarchizes the plurality of time series data included in the second time series data set according to the determined temporal relationship, and outputs the hierarchization result; a visualization unit (160) that generates a visualized diagram that visualizes the plurality of time series data included in the second time series data set based on the calculated waveform semantic relationship and the output hierarchization result; and a recalculation unit (170) that receives additional time series data and performs recalculation to add the received additional time series data to the visualized diagram.

[0013] (Data entry section) The data input unit 110 acquires a time-series data set (first time-series data set) D1 that includes multiple time-series data that are candidates for leading indicators. From among the candidates for leading indicators, leading indicators or time-series data to be used for predicting future indicators are selected. An example of a prediction is a future indicator such as the number of air conditioners shipped.

[0014] Here, the purpose of the time series data processing device 100 and the time series data processing system of the present disclosure will be explained in more detail. To this end, let us consider a case where apple consumption is used as an explanatory variable when predicting the number of air conditioner shipments. Even if this prediction were successful, it would be difficult to find a causal relationship between "apple consumption" and "number of air conditioner shipments" from domain knowledge. Therefore, the prediction results in this case may not be convincing to the user of the time series data processing device 100, potentially causing distrust among the user.

[0015] In contrast, let's assume that the number of new building construction starts is used to explain the forecast of the number of air conditioner shipments. In this case, domain knowledge indicates that there is a causal relationship: "Increase in the number of buildings → Increase in demand for new air conditioners to be installed in buildings → Increase in the number of air conditioner shipments." Therefore, the forecast results in this case will be convincing to the user.

[0016] It is difficult to select explanatory variables that are convincing to users using statistical judgment alone, and ultimately manual judgment is essential. However, for example, there are hundreds of thousands of types of economic indicator data used in product demand forecasting, and it is extremely difficult to manually select explanatory variables that are valid from such a large amount of data, both in terms of domain knowledge and statistics.

[0017] The time-series data processing device 100 and the time-series data processing system of the present disclosure aim to improve the method of selecting explanatory variables, thereby giving users a sense of satisfaction and reducing distrust in the predictions. The user's sense of satisfaction means the meaning of the causal relationships between multiple indicators.

[0018] Returning to the description of the data input unit 110, the time series dataset D1 includes multiple time series datasets, each of which is assigned a name indicating what kind of data it is. The time series of the multiple time series datasets included in the time series dataset D1 are all in the same unit. For example, the multiple time series datasets in the time series dataset D1 are all in the same unit, such as all monthly data or all yearly data. The data input unit 110 supplies the acquired time series dataset D1 to the preprocessing unit 120.

[0019] (Preprocessing section) The preprocessing unit 120 is a functional unit that performs data preprocessing on the time-series dataset D1 acquired by the data input unit 110 to generate a time-series dataset D2 and outputs the generated time-series dataset D2 to the grouping unit 130. That is, the preprocessing unit 120 converts the relationships between multiple time-series data included in the time-series dataset D1 into a calculable format, and generates a second time-series dataset including the converted multiple time-series data. Specifically, the data preprocessing includes processing of missing values, data standardization, etc.

[0020] (Grouping section) The grouping unit 130 is a functional unit that classifies the time-series data set D2 into multiple groups according to the properties of each data, generates a representative value for each group, assigns that information (representative value) to the group, and outputs the group D3 after the representative value has been assigned to the relationship calculation unit 140.

[0021] More specifically, the grouping unit 130 first performs clustering on the time-series data set D2 obtained from the preprocessing unit 120 based on the similarity of waveforms or names. The resulting clusters are treated as groups, and data processing is performed on a group-by-group basis thereafter. Next, in order to treat each group as time-series data, a representative value that reflects the characteristics of each group is generated.

[0022] In the first embodiment, the time series data set D2 is classified as follows. First, the cross-correlation is calculated with the time series of each data aligned according to the following formula (1). Next, data whose correlation coefficient is equal to or greater than a certain value (arbitrary, D3-N1) is classified into the same group. The group contains information on the name list (D3-1) of the data included in that group. TIFF0007814628000001.tif15166 Here, index i is date and time t, where t is, for example, January 1st, February 1st, March 1st, . . . , December 1st.

[0023] The representative value is generated as follows: The period for the representative value of each group is set to the oldest point in time among the data included in that group, from the newest point in time, and this is set as period T1. The average value of the data included in that group at each point in time within period T1 is set as the representative value of that group at each point in time. In this way, a representative value that reflects the characteristics of the data included in each group can be generated. Group D3 is set to include information (D3-2) about this representative value.

[0024] FIG. 2 is a diagram illustrating an outline of processing by the grouping unit 130. As illustrated in FIG. 2, the time-series dataset D2 includes multiple time-series datasets, such as a time-series dataset of "iron resource production volume," a time-series dataset of "electricity production volume," a time-series dataset of "number of construction starts," and a time-series dataset of "food export volume." The grouping unit 130 groups the multiple time-series datasets included in the time-series dataset D2 according to the properties of the data. FIG. 2 illustrates how the time-series datasets of "iron resource production volume" and "electricity production volume" are classified into group 1, and the time-series datasets of "number of construction starts" and "food export volume" are classified into group 2. In this way, the time-series datasets are classified into groups, and each group after classification is referred to as group D3.

[0025] (Relationship Calculation Unit) The relationship calculation unit 140 is a functional unit that calculates the relationship between groups (waveform semantic relationship) D4 for the groups D3 generated by the grouping unit 130 and outputs the relationship D4 and information about the groups D3 to the hierarchical unit 150. More specifically, the relationship between multiple groups D3 is calculated using a representative value for each group. The relationship between groups includes information about the waveform similarity (D4-1) between the groups. The waveform semantic relationship refers to a waveform relationship or a semantic relationship. The waveform relationship is an index of the degree of similarity between waveforms of data. The semantic relationship is the depth of the association between data, which can be defined, for example, by domain knowledge. As a specific example of a semantic relationship, for example, there is a case where domain knowledge reveals that there is a clear relationship between data on air conditioner demand and data on average temperature, even if the waveforms are not similar.

[0026] In the first embodiment, the relationship between groups is calculated as follows. Since the representative value of each group can be treated as time-series data, the relationship is defined by the magnitude of the cross-correlation coefficient between the data, similar to the processing in the grouping unit 130. However, when calculating the cross-correlation in the relationship calculation unit, the calculation is performed using data in which the representative values ​​of the groups are shifted forward or backward by one month, up to ±N (any integer; for example, 12) months, and the largest correlation coefficient among them is taken as the waveform similarity (D4-1) between the groups. Furthermore, the amount of time shift at this time is taken as the temporal precedence relationship (D4-2) between the groups.

[0027] 3 is a diagram showing an overview of the relationship calculation performed by the relationship calculation unit 140. As shown in FIG. 3, relationships D4 between multiple groups D3 are calculated, and groups with strong relationships are connected by lines according to the calculation results. In FIG. 3, group 1 has strong relationships with groups 2 and 3, and group 1 is connected by lines to groups 2 and 3. Group 2 has strong relationships with groups 4 and 5 in addition to group 1, and group 2 is also connected by lines to groups 4 and 5. Group 4 has a strong relationship with group 6 in addition to group 2, and group 4 is also connected by lines to group 6.

[0028] (layering section) The layering unit 150 is a functional unit that determines the temporal relationships between the multiple time series data included in the time series data set D2, layers the multiple time series data included in the time series data set D2 according to the determined temporal relationships, and outputs the layering results.

[0029] The layering unit 150 may process the multiple groups D3 generated by the grouping unit 130. In this case, the layering unit 150 determines the temporal relationship between the multiple groups D3, layers the multiple groups D3 in accordance with the determined temporal relationship, and outputs the layering result.

[0030] This will be described in more detail below in relation to a case where multiple groups D3 are processed by the layering unit 150. In order to clarify the causal relationships between the multiple groups D3, the layering unit 150 determines the temporal relationships between the groups D3 based on a statistical analysis of the process order of various activities or month-shift correlations, and assigns information (D3-3) about the temporal relationships to the groups D3.

[0031] Fig. 4 is a diagram showing an overview of an example of hierarchization performed by the hierarchization unit 150. As shown in Fig. 4, the hierarchization unit 150 hierarchically rearranges group D3 connected by the relationship calculation unit 140 according to the chronological order. The hierarchical example in Fig. 4 shows that group 2 precedes group 4 and group 5 in time, and group 1 precedes group 2 and group 3 in time.

[0032] As an example, in the first embodiment, the hierarchy of groups (temporal relationships) is generated as follows. That is, the time series data are shifted in time so as to maximize the correlation coefficient, and the temporal precedence relationship between the data is determined based on the magnitude of the shift. This point will be described with reference to FIG. 5. FIG. 5 is a diagram illustrating an example of a month-shifted correlation, where FIG. 5A shows the original waveform and FIG. 5B shows the waveform when one of the waveforms is shifted by one month. As shown in FIG. 5, assume that there is a relationship between some time series data X (waveform represented by a thick line) and another time series data Y (waveform represented by a thin line) such that when the time series data Y is shifted forward by one month, the cross-correlation coefficient between the time series data X and Y becomes the highest. This means that the time series data Y precedes the time series data X by one month. Therefore, a temporal precedence relationship is obtained in which the time series data Y precedes the time series data X by one month. The layering unit 150 also grasps such temporal precedence relationships for other data, determines temporal context relationships from the grasped temporal precedence relationships, layers multiple groups D3 according to the determined temporal context relationships, and outputs the layering results. Similarly, when the time series dataset D2 is the processing target, the layering unit 150 determines temporal context relationships between multiple time series data included in the time series dataset D2, layers multiple time series data included in the time series dataset D2 according to the determined temporal context relationships, and outputs the layering results.

[0033] In another embodiment of layering, for example, a temporal context of words, such as demand → production → sales, is set in advance according to domain knowledge. The domain knowledge is defined and performed by a user input. The preset domain knowledge includes multiple words and definitions of the temporal context between the multiple words. The layering unit 150 acquires the set domain knowledge. The user input for the pre-setting may be acquired directly by the layering unit 150 or indirectly via the data input unit 110. Based on the acquired temporal context, the layering unit 150 hierarchizes data whose names include those words. FIG. 6 is a diagram showing a specific example of layering according to a preset temporal context. The hierarchizing unit 150 acquires a preset temporal sequence of demand → production → sales, and stratifies the groups of "private car demand," "private car production," "private car sales," "metal demand," "semiconductor demand," "PC production," and "home appliance sales" according to this temporal sequence. Because the groups of "private car demand," "metal demand," and "semiconductor demand" contain the term "demand," the hierarchizing unit 150 classifies the groups of "private car demand," "metal demand," and "semiconductor demand" as belonging to the first hierarchical layer. Because the groups of "private car production" and "PC production" contain the term "production," the hierarchizing unit 150 classifies the groups of "private car production" and "PC production" as belonging to the second hierarchical layer. Since the groups "Personal vehicle sales" and "Home appliance sales" contain the term "sales," the hierarchical classification unit 150 classifies the groups "Personal vehicle sales" and "Home appliance sales" as belonging to the third hierarchical level.

[0034] In addition to the above, the stratification unit 150 may perform stratification using the Granger causality test method.

[0035] (Visualization part) The visualization unit 160 is a functional unit that uses information on groups D3 (including a list D3-1 of the names of the data contained in each group and information D3-3 on the process order of each group) and (semantic) relationships D4 between groups (including waveform similarities D4-1 between groups and temporal precedence relationships D4-2 between groups) to generate a visualization diagram that visualizes the relationships between groups in a way that makes it easier for humans to understand, and outputs the generated visualization diagram.

[0036] In the first embodiment, the relationships between groups are visualized, for example, as follows: The relationships between groups are visualized in a graph format, with each group (D3) as a vertex and the relationships between groups (D4) as edges.

[0037] First, a hierarchical structure similar to the hierarchy (D3-3) defined by the hierarchy unit 150 is prepared.

[0038] Next, based on the list of data names contained in each group (D3-1), each group is assigned to a hierarchy. Each group is visualized as a vertex.

[0039] Next, the relationships between the groups are visualized based on the waveform similarity information between the groups (D4-1). When visualizing, the groups, which are vertices, are connected by edges with different properties, such as line thickness, line color, or line type (solid or dashed), depending on the strength of the relationship.

[0040] The list of data names (D3-1) included in each group can be easily viewed by, for example, clicking on a group of vertices to display it in list form. Such operations and displays are performed via an input / output device (not shown), and the visualization unit 160 acquires operation instructions input via the input device and performs display control to display them on an output device.

[0041] (Recalculation part) After the series of steps from the data input unit 110 to the visualization unit 160 has been executed once, the recalculation unit 170 accepts additional time series data D1-2 and performs processing to output a visualization diagram in which the accepted additional time series data D1-2 has been added to the time series data set D1.

[0042] In the first embodiment, recalculation when additional time series data D1-2 is added is performed as follows: The additional time series data D1-2 is stored in the storage device 300, and the recalculation unit 170 acquires the additional time series data D1-2 from the storage device 300.

[0043] Here, the additional time series data D1-2 is assumed to be a single piece of time series data. When adding multiple pieces of time series data, the additional data are added to the generated visualization by performing the following itemized process for each piece of additional data. The recalculation unit 170 supplies the additional time series data D1-2 to the preprocessing unit 120. The preprocessing unit 120 performs preprocessing on the additional time series data D1-2. The grouping unit 130 calculates the cross-correlation coefficient between the additional time-series data D1-2 after preprocessing and the multiple groups D3 that have already been generated. The cross-correlation coefficient may be calculated using the same method as described above. If there is any cross-correlation coefficient equal to or greater than the arbitrary constant (D3-N1) set by the grouping unit 130, the grouping unit 130 selects the group with the highest cross-correlation coefficient and adds the additional time-series data D1-2 to the selected group. That is, the grouping unit 130 adds the name of the additional data to the list D3-1 of names of data included in the group. Completion of this process completes the recalculation. If there is no correlation coefficient equal to or greater than the arbitrary constant (D3-N1) set by the grouping unit 130, the grouping unit 130 groups the additional time series data D1-2 into a single group. The relationship calculation unit 140 then calculates the relationship between the single group of additional time series data D1-2 and another group D3. Based on the results, the visualization unit 160 updates the visualization diagram and completes the recalculation.

[0044] Note that instead of the method configured by the above items, the recalculation process may be performed by another method. Specifically, the other method is as follows: The recalculation unit 170 merges the additional time series data D1-2 with the original time series data set D1 to generate a time series data set D1-3. The recalculation unit 170 supplies the generated time series data set D1-3 to the preprocessing unit 120. A visualization diagram is newly generated from scratch by performing processing on the time series data set D1-3 from the preprocessing unit 120 to the visualization unit 160. The selection of these methods is made based on user input.

[0045] (Hardware) Next, with reference to Figures 7A and 7B, an example of the hardware configuration of the time series data processing device 100 will be described. Each function of the time series data processing device 100 is realized by a processing circuitry. The processing circuitry may be a dedicated processing circuit 400 as shown in Figure 7A, or a processor 500 that executes a program stored in a memory 600 as shown in Figure 7B.

[0046] When the processing circuitry is a dedicated processing circuit 400, the dedicated processing circuit 400 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 time series data processing device 100 may be realized by a plurality of separate processing circuits, or the functions of the time series data processing device 100 may be realized together in a single processing circuit.

[0047] When the processing circuitry is a processor 500, the functions of the time-series data processing device 100 are realized by software, firmware, or a combination of software and firmware. The software and firmware are written as programs and stored in a memory 600. The processor 500 realizes the functions of the time-series data processing device 100 by reading and executing the programs stored in the memory 600. Examples of the memory 600 include non-volatile or volatile semiconductor memories such as random access memory (RAM), read-only memory (ROM), flash memory, erasable programmable read-only memory (EPROM), and electrically erasable programmable read-only memory (EEPROM), as well as magnetic disks, flexible disks, optical disks, compact disks, minidisks, and DVDs. The memory 600 may be realized as the same device as the storage device 200 or the storage device 300.

[0048] It is also possible to realize some of the functions of the time series data processing device 100 with dedicated hardware, and other functions with software or firmware. In this way, the processing circuit can realize the functions of the time series data processing device 100 with hardware, software, firmware, or a combination of these.

[0049] <Operation> Next, the operation of the time-series data processing device 100 will be described with reference to FIG.

[0050] (Step ST1) First, in step ST1, the data input unit 110 acquires a time-series data set D1 including candidates for explanatory variables to be used for prediction.

[0051] (Step ST2) Next, in step ST2, the preprocessing unit 120 performs preprocessing on the time-series data set D1 acquired by the data input unit 110. The preprocessing unit 120 supplies the time-series data set D2 obtained by performing preprocessing on the time-series data set D1 to the grouping unit 130.

[0052] (Step ST3) Next, in step ST3, the grouping unit 130 classifies and groups the acquired time-series data set D2 according to the similarity of the properties of each piece of data.

[0053] (Step ST4) Next, in step ST4, the relationship calculation section 140 calculates the relationship between the groups and also obtains the temporal relationship between the groups by comparing the waveforms.

[0054] (Step ST5) Next, in step ST5, the hierarchizing unit 150 determines the process order of various activities or the monthly shift correlation based on statistical analysis, calculates the temporal relationship between the activities, and hierarchizes the groups based on the calculated temporal relationship.

[0055] (Step ST6) Next, in step ST6, the visualization unit 160 hierarchically visualizes the relationships between the groups obtained in step ST4 based on the temporal context obtained in step ST5.

[0056] (Step ST7) After performing a series of processes from steps ST1 to ST6, the recalculation unit 170 recalculates the relationships in order to add the additional time-series data to the existing groups and visualize them. When adding time-series data, the recalculation of the relationships can be performed while reducing time costs by comparing the additional data with the existing groups.

[0057] <Effects> When predicting indicators of economic activity, manual judgment is ultimately essential to select explanatory variables that are convincing to users. However, it is extremely difficult to manually select explanatory variables that are both domain knowledge-based and statistically appropriate from a large amount of data.

[0058] The visualized diagram of relationships output by the time-series data processing device 100 of the present disclosure makes it possible to extract several candidates as statistically valid explanatory variables from a large amount of data. By extracting multiple candidates in this manner, it is possible to narrow down the number of candidates that humans must select based on their domain knowledge. This makes it easier for the humans making the predictions to select explanatory variables themselves. Therefore, it is possible to incorporate human domain knowledge while ensuring the statistical usefulness of the explanatory variables used in the predictions, thereby reducing distrust in predictions obtained from those explanatory variables and improving the sense of convincingness of the predictions.

[0059] <Example> A more specific embodiment of the time series data processing device 100 will be described below. A visualized diagram of relationships output by the time series data processing device 100 is, for example, as shown in FIG. 9. In FIG. 9, each vertex represents a group of similar time series data, and the edges connecting the vertices represent the strength of the relationships between the groups. Furthermore, the multiple hierarchies in FIG. 9 are hierarchies distinguished according to temporal relationships, and it is shown that groups in higher hierarchies in the diagram precede groups below them in time.

[0060] As an example, consider the case where future trends are predicted for data on "number of air conditioner shipments" included in group 2 in FIG.

[0061] (1st step) From the information on the relationships between data in Figure 9, it can be seen that Group 2, to which the data "Number of air conditioner shipments" belongs, has a strong relationship with Groups 1, 4, and 5. This makes it possible to narrow down the candidates by determining that the data that is useful as an explanatory variable to use when predicting the data "Number of air conditioner shipments" is data contained in Groups 1, 4, or 5.

[0062] (2nd process) From the hierarchical information in Figure 9, it can be seen that of the groups (1, 4, and 5) narrowed down in the first step, only Group 1 precedes in time Group 2, to which the data "Number of air conditioners shipped" belongs. This makes it possible to narrow down the candidates by assuming that the data contained in Group 1 are useful as explanatory variables to use when predicting the data "Number of air conditioners shipped." When predicting certain data Y, if there is another data X that precedes that data in time, then data X will be a useful explanatory variable (leading indicator) in predicting data Y.

[0063] (3rd step) It can be assumed that the information on the groups in Figure 9 indicates that the data "number of new building construction starts," "apple consumption," and "mackerel catch" belong to the group of explanatory variable candidates narrowed down in the second step. These data "number of new building construction starts," "apple consumption," and "mackerel catch" are presented to the user as final explanatory variable candidates.

[0064] (4th step) The user selects the data that is most convincing when used for prediction from the above-mentioned candidate explanatory variable data "number of new building construction starts," "apple consumption," and "mackerel catch volume." For example, there is an intuitively easy-to-understand causal relationship between the data "number of air conditioner shipments" and the data "number of new building construction starts," because "increase in the number of buildings → increase in demand for air conditioners to be newly installed in buildings → increase in number of air conditioner shipments." Therefore, the user can easily select "number of new building construction starts" as the explanatory variable to use for predicting "number of air conditioner shipments." In this way, by making a prediction using data with an intuitively easy-to-understand causal relationship, the user can be convinced of the prediction results.

[0065] The number of factors appearing in the visualization diagram can be reduced by grouping the explanatory variables using the grouping unit 130. As a result, it becomes possible to gradually narrow down the candidates for explanatory variables, thereby reducing the effort required to select and discard explanatory variables.

[0066] By visualizing the relationships between candidate explanatory variables, it becomes possible to select explanatory variables without relying on human ability, something that previously relied on the domain knowledge and tacit knowledge of experts.

[0067] When adding data for candidate explanatory variables or data to be predicted, the explanatory variables are grouped by the grouping unit 130, and the added data is assigned to existing groups according to the process of the recalculation unit 170. This eliminates the need to repeat grouping, relationship calculation, and visualization, thereby reducing the time cost required to re-output the visualization diagram.

[0068] It is possible to combine the embodiments, and to modify or omit each embodiment as appropriate. [Industrial Applicability]

[0069] The time-series data processing device of the present disclosure can be used as a device for predicting data related to indicators such as the number of air conditioner shipments. [Explanation of symbols]

[0070] 100 time series data processing device, 110 data input unit, 120 preprocessing unit, 130 grouping unit, 140 relationship calculation unit, 150 hierarchical unit, 160 visualization unit, 170 recalculation unit, 200 storage device, 300 storage device, 400 processing circuit, 500 processor, 600 memory.

Claims

1. a data input unit that acquires a first time series data set including a plurality of time series data to be candidates for explanatory variables; a pre-processing unit that converts the relationships between the plurality of time series data included in the acquired first time series data set into a calculable format and generates a second time series data set including the converted plurality of time series data; a relationship calculation unit that calculates a similarity between a plurality of time series data included in the second time series data set, and calculates the calculated similarity or a relevance obtained as a user input between the plurality of time series data included in the second time series data set as a waveform semantic relationship between the plurality of time series data included in the second time series data set; a hierarchizing unit that calculates a temporal relationship between a plurality of time series data included in the second time series data set, hierarchizes the plurality of time series data included in the second time series data set according to the calculated temporal relationship, and outputs the hierarchization result; a visualization unit that generates a visualization diagram that visualizes the plurality of time series data included in the second time series data set based on the calculated waveform semantic relationships and the output hierarchical structure results; and Equipped with the visualized diagram is a graph in which each time series data item included in the second time series data set is represented as a vertex and a waveform semantic relationship between the time series data items is represented as an edge. Time series data processing device.

2. the hierarchizing unit determines a temporal relationship between a plurality of time series data included in the second time series data set according to domain knowledge defined by a user input; 2. The time series data processing device according to claim 1.

3. The domain knowledge includes a plurality of words and a definition of a temporal context between the plurality of words.

3. A time series data processing device according to claim 2.

4. the layering unit calculates a shift width by performing time shifting for the plurality of time series data included in the second time series data set so that a correlation coefficient between the plurality of time series data becomes the highest, and calculates a temporal relationship between the plurality of time series data included in the second time series data set from the calculated shift width; 2. The time series data processing device according to claim 1.

5. the similarity between the plurality of time series data included in the second time series data set is a cross-correlation between the plurality of time series data included in the second time series data set; 2. The time series data processing device according to claim 1.

6. the edges have different line thicknesses, line colors, or line types (solid lines or dashed lines) depending on waveform semantic relationships between the plurality of time series data included in the second time series data set; 2. The time series data processing device according to claim 1.

7. a grouping unit that classifies a plurality of time series data included in the second time series data set into groups according to the properties of the data and generates a representative value for each group; The time series data processing device according to claim 1 , further comprising:

8. the grouping unit groups the plurality of time series data included in the second time series data set based on a similarity between waveforms of the plurality of time series data included in the second time series data set; 8. The time series data processing device according to claim 7.

9. the grouping unit generates a representative value for each group that reflects characteristics of the data included in each group; 8. The time series data processing device according to claim 7.

10. a recalculation unit that receives additional time series data and performs recalculation to add the received additional time series data to the visualization diagram; The time series data processing device according to any one of claims 1 to 9, further comprising:

11. the recalculation unit determines a representative value of the received additional time series data, calculates a correlation coefficient between the determined representative value of the received additional time series data and the generated representative value of each group, and assigns the received additional time series data to the group for which the calculated correlation coefficient is largest. The time series data processing device according to claim 10.

12. the recalculation unit determines a representative value of the received additional time series data, calculates a correlation coefficient between the determined representative value of the received additional time series data and a representative value of each generated group, and if any of the calculated correlation coefficients is less than a predetermined threshold, generates a new group including the received additional time series data, and calculates a waveform semantic relationship and a temporal context relationship between the generated new group and other groups. The time series data processing device according to claim 10.

13. A time series data processing method performed by a time series data processing device including a data input unit, a preprocessing unit, a relationship calculation unit, a hierarchical unit, and a visualization unit, comprising: the data input unit acquiring a first time series data set including a plurality of time series data to be candidates for explanatory variables; a preprocessing unit converting the relationships between the plurality of time series data included in the acquired first time series data set into a calculable format, and generating a second time series data set including the converted plurality of time series data; the relationship calculation unit calculates a similarity between a plurality of time series data included in the second time series data set, and calculates the calculated similarity or a relevance obtained as a user input between the plurality of time series data included in the second time series data set as a waveform semantic relationship between the plurality of time series data included in the second time series data set; the layering unit determining a temporal relationship between a plurality of time series data included in the second time series data set, layering the plurality of time series data included in the second time series data set according to the determined temporal relationship, and outputting the layering result; a step in which the visualization unit generates a visualization diagram that visualizes a plurality of time series data included in the second time series data set based on the calculated waveform semantic relationship and the output hierarchical result, the visualization diagram being a graph in which each time series data item included in the second time series data set is represented as a vertex and the waveform semantic relationship between the time series data items is represented as an edge; A time series data processing method comprising:

Images