Data analysis device and data analysis method

Abstract

To provide a data analysis device and data analysis method that can analyze characteristics of time series data using a difference of input / output characteristics of neural networks.SOLUTION: A data analysis device (100) for analyzing characteristics of time series data includes: acquisition sections (41, 43) that acquire time series data to be analyzed, time series prediction data obtained by inputting the time series data to be analyzed to a neural network used for time series prediction, and information on input / output characteristics of the neural network; a data processing section (41) that compares the time series prediction data with the time series data to be analyzed; and an analysis section (43) that analyzes characteristics of the time series data to be analyzed on the basis of a result of comparing time series prediction data of each of a plurality of neural networks with the time series data to be analyzed, and the input / output characteristics of each of the neural networks.SELECTED DRAWING: Figure 2

Description

[Technical field]

[0001] The present invention relates to a data analysis device and a data analysis method for analyzing characteristics of time-series data. [Background technology]

[0002] In recent years, time series prediction using neural networks has been utilized in various fields. One aspect of neural networks used for time series prediction is, for example, a recurrent neural network (RNN). As recurrent neural networks, various neural networks are known, such as a fully recurrent neural network, a deep recurrent neural network, an independent recurrent neural network, a long-short-term memory network (LSTM), and a gated recurrent unit (GRU). [Prior art documents] [Patent documents]

[0003] [Patent Document 1] JP 2020-067948 A [Patent Document 2] JP 2020-024689 A Summary of the Invention [Problem to be solved by the invention]

[0004] It is known that each of these neural networks has different input / output characteristics. In other words, even when the same time series data is input, the time series prediction data output may differ depending on the type of neural network. In addition, even when the neural networks are of the same type, the input / output characteristics may change if the number of layers of the neural network or the number of input data is different. For example, when a certain time series data is input to each neural network, some neural networks may have a high reproducibility of the output time series prediction data, while others may have a low reproducibility, depending on the input / output characteristics.

[0005] Here, events that progress chronologically often reflect the thoughts and ideas of the subject of the events. For example, the way in which steps are played in video games such as Tetris, or in matches such as Go or Shogi, reflects the thoughts and ideas of the player or shogi player (hereinafter collectively referred to as "player"). Specifically, the content of a play varies from deeply thought out plays to shallow plays, and even plays with an offensive stance to plays with a defensive stance. In other words, it is believed that the thinking level and character of a player can be estimated from the chronological data of the plays made by that player.

[0006] An object of the present invention is to provide a data analysis device and a data analysis method capable of analyzing the characteristics of time-series data by utilizing the differences in the input / output characteristics of the above-mentioned neural networks. [Means for solving the problem]

[0007] In order to solve the above problem, according to one aspect of the present invention, there is provided a data analysis device for analyzing characteristics of time series data, the data analysis device including: an acquisition unit that acquires information regarding the time series data of the analysis target, time series prediction data obtained by inputting the time series data of the analysis target into a neural network used for time series prediction, and input / output characteristics of the neural network; a data processing unit that compares the time series prediction data with the time series data of the analysis target; and an analysis unit that analyzes the characteristics of the time series data of the analysis target based on the results of the comparison of the time series prediction data for each of a plurality of neural networks with the time series data of the analysis target and the input / output characteristics of each neural network.

[0008] In addition, in order to solve the above problem, according to another aspect of the present invention, there is provided a data analysis method for analyzing characteristics of time series data, the data analysis method including the steps of acquiring information regarding the time series data to be analyzed, time series prediction data obtained by inputting the time series data to be analyzed into a neural network used for time series prediction, and input / output characteristics of the neural network, comparing the time series prediction data with the time series data to be analyzed, and analyzing the characteristics of the time series data to be analyzed based on the results of comparing the time series prediction data for each of a plurality of neural networks with the time series data to be analyzed, and the input / output characteristics of each neural network. Effect of the Invention

[0009] As described above, according to the present invention, the characteristics of time-series data can be analyzed by utilizing the differences in input / output characteristics of a recurrent neural network. [Brief description of the drawings]

[0010] [Figure 1] 1 is a schematic diagram illustrating an example of a configuration of a data analysis device according to an embodiment of the present invention. [Diagram 2] FIG. 1 is a schematic diagram showing a time series prediction process using an RNN. [Diagram 3] 11 is a flowchart illustrating an example of a data analysis method according to the embodiment. [Figure 4] 13 is a flowchart illustrating an example of a data analysis method according to another embodiment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0011] Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings. In this specification and the drawings, components having substantially the same functional configurations are designated by the same reference numerals, and duplicated explanations will be omitted.

[0012] <1. Example of data analysis device configuration> First, a configuration example of a data analysis device according to an embodiment of the present invention will be described.

[0013] 1 is a schematic diagram showing an example of the configuration of a data analysis device 100 according to this embodiment. The data analysis device 100 includes an input unit 10, a display unit 20, a storage unit 30, and a processor 40.

[0014] There is no particular limitation on the time-series data to be analyzed by data analysis apparatus 100. The time-series data to be analyzed may be data on the procedure of a predetermined operation, or may be time-series image data capturing a change in a predetermined state.

[0015] The input unit 10 has a function of reading data to be input to the arithmetic processing device 40. The input unit 10 is used to input at least time-series data to be analyzed. For example, the input unit 10 may be a communication interface for receiving data transmitted from other computer devices, an interface to which a data storage medium such as a Universal Serial Bus (USB) flash memory is connected, or a data playback device into which a data storage medium such as a Compact Disc (CD) or a Digital Versatile Disc (DVD) is inserted.

[0016] The input unit 10 may be a detection device that generates image data of the object to be photographed and transmits it to the arithmetic processing device 40 as time-series data. Examples of such detection devices include an imaging camera or LiDAR (Light Detection and Ranging or Light Imaging Detection and Ranging). The imaging camera includes an imaging element such as a CCD (Charge Coupled Device) or a CMOS (Complementary Metal Oxide Semiconductor) and generates image data of the object to be photographed. The LiDAR generates image data reflecting the external state of the object to be measured by irradiating the object with a laser light while scanning it and observing the scattering and reflected light of the laser light.

[0017] The input unit 10 may also have a function of accepting an operation input by a user and transmitting an operation signal to the arithmetic processing device 40. For example, the input unit 10 may be configured to include at least one of a mouse, a keyboard, a touch panel, a tablet computer, a smartphone, an operation button, a switch, and the like.

[0018] The display unit 20 is driven based on a drive signal output from the arithmetic processing device 40, and performs a predetermined display of at least the analysis results of time-series data executed by the arithmetic processing device 40. In addition, the display unit 20 displays an image of the operation screen of the data analysis device 100. For example, the display unit 20 is configured to include an optical panel such as a liquid crystal panel. Note that the input unit 10 and the display unit 20 may be integrated to form a touch panel.

[0019] The storage unit 30 is configured with at least one storage medium such as a magnetic disk, a hard disk drive (HDD), a compact disc (CD)-ROM, a digital versatile disc (DVD), a solid state drive (SSD), a universal serial bus (USB) flash, a storage device, etc. The storage unit 30 stores a plurality of recurrent neural networks (hereinafter sometimes referred to as "RNN") used for time series prediction.

[0020] The RNNs stored in the storage unit 30 include, for example, one or more of a full recurrent neural network, a deep recurrent neural network, an independent recurrent neural network, a long short-term memory network (LSTM), a gated recurrent unit (GRU), etc. The RNNs stored in the storage unit 30 may include multiple RNNs of the same type that differ in at least one of the number of layers or the number of input data.

[0021] In this embodiment, the multiple RNNs stored in the storage unit 30 are RNNs that have previously learned teacher data. The teacher data is one or more time series data related to the time series data to be analyzed. For example, if the time series data to be analyzed is the content of playing Tetris, the teacher data is also time series data of the content of playing Tetris. Each RNN has repeatedly learned the teacher data until time series prediction data for a predetermined time series of input data becomes stable a predetermined number of times or for a predetermined time.

[0022] The arithmetic processing device 40 includes a processor such as a CPU (Central Processing Unit). The arithmetic processing device 40 may include a GPU (Graphic Processing Unit) in addition to the CPU. The arithmetic processing device 40 includes a communication interface (not shown) for transmitting and receiving signals to and from the input unit 10 or the display unit 20.

[0023] In this embodiment, the arithmetic processing device 40 includes a data processing unit 41, an analysis unit 43, a display control unit 45, and a memory 47. Among these, the memory 47 stores programs executed by the processor, various calculation parameters used in the arithmetic processing, acquired data, calculation results, etc. Moreover, the data processing unit 41, the analysis unit 43, and the display control unit 45 are functions realized by the processor executing the programs.

[0024] The data processing unit 41 executes a process of acquiring time series prediction data obtained by inputting the time series data of the analysis target to the multiple RNNs stored in the storage unit 30, and the time series data of the analysis target, and comparing the acquired time series prediction data with the time series data of the analysis target. In this embodiment, the data processing unit 41 also functions as an acquisition unit. In this embodiment, the data processing unit 41 calculates the reproducibility of the time series prediction data obtained by each RNN with respect to the time series data of the analysis target, and stores the result in the memory 47.

[0025] 2 is a schematic diagram showing how time series data ipt_ts (x0, x1, x2, x3... xn) to be analyzed is input to an RNN, and time series prediction data opt_ts (y0, y1, y2, y3... yn) is output. The number of layers Dep_RNN of the RNN is indicated by "m". The data processing unit 41 compares the time series prediction data opt_ts (y0, y1, y2, y3... yn) obtained from each RNN with the time series data ipt_ts (x0, x1, x2, x3... xn) to be analyzed, and calculates the degree of reproducibility.

[0026] The reproducibility may be defined appropriately according to the type of time series data and the contents of the time series to be analyzed. For example, if the predicted value x*(t+1) (=y(t+1)) at time t+1 output when time series data x(tm), ..., x(t) up to a certain time t is input to an RNN is expressed by the following formula (1), the reproducibility may be the error function Err(X) between the output predicted value x*(t+1) and the actual data x(t) at time t+1, as shown in the following formula (2). Note that the input data x(t) is not limited to a single value and may be image data.

[0027]

number

[0028]

number

[0029] Furthermore, if the example error function Err(X) is expressed at a coarse level, for example as the recall Err(x(t-1), x(t)), then the recall can also be defined as the expected value for the input data x(tm), ..., x(t-2).

[0030] The analysis unit 43 analyzes the characteristics of the time series data to be analyzed based on the result of the comparison by the data processing unit 41 between the time series prediction data for each of the multiple RNNs and the time series data to be analyzed, and the characteristics of each RNN. Specifically, the analysis unit 43 acquires information on the comparison result for the multiple RNNs by the data processing unit 41 and information on the input characteristics indicating the input / output characteristics of each RNN, and analyzes the characteristics of the time series data to be analyzed based on the acquired information. In this embodiment, the analysis unit 43 also functions as an acquisition unit. As already mentioned, the input / output characteristics of RNNs differ depending on the type, the number of layers, or the number of input data.

[0031] In the above example, the fully recurrent neural network has input / output characteristics that are strong in simple pattern memory and prediction, focusing on the shortest-term memory. The deep recurrent neural network has input / output characteristics that are strong in complex pattern memory and prediction compared to general RNNs. The independent recurrent neural network has input / output characteristics that are strong in long-term memory and prediction because of its higher learning effect compared to the fully recurrent neural network. The long short-term memory network (LSTM) has input / output characteristics that are strong in time series with segmentation such as natural language. The gated recurrent unit (GRU) has input / output characteristics that are strong in time series with segmentation like LSTM, but has a characteristic that it tends to be less prone to overfitting than LSTM because it has fewer parameters.

[0032] Furthermore, even for the same type of RNN, the more layers there are, the stronger the input / output characteristics are in terms of memorizing and predicting complex patterns.Also, even for the same type of RNN, the more input data there is, the more distant the predictions can be made.

[0033] Therefore, when the reproducibility of predicted data from a certain RNN is high, the time series data to be analyzed can be evaluated as time series data having the input / output characteristics of the RNN. On the other hand, when the reproducibility of predicted data from a certain RNN is low, the time series data to be analyzed can be evaluated as time series data not having the input / output characteristics of the RNN.

[0034] For example, if the time series data to be analyzed is data on the content of Tetris play, and the reproducibility of the time series prediction data output from an RNN with deep layers (a large number of layers) is high, while the reproducibility of the time series prediction data output from an RNN with shallow layers (a small number of layers) is low, the content of the play can be characterized as a play that was deeply thought out. Conversely, if the reproducibility is high even for an RNN with shallow layers and a short number of input data, the content of the play can be characterized as a play that was shallowly thought out.

[0035] In addition, in Othello, if you control the four corners, you will ultimately get a very good result, so it is effective to go for the four corners even if it seems to be a short-term disadvantageous play. Although such plays are strategic, they can be predicted by a simple RNN with shallow layers and a short amount of input data, depending on the training data. This is because predictions can be made just by looking at the board in front of you, and there is no need for complex pattern recognition or long-term memory. Conversely, if a time series requires complex pattern recognition or long-term memory, it is possible to analyze the characteristics of that time series by comparing the reproducibility of the predicted data of the time series output from each RNN.

[0036] The display control unit 45 generates an image signal to be displayed on the display unit 20, and controls the display on the display unit 20. The display control unit 45 causes at least the analysis results by the analysis unit 43 to be displayed on the display unit 20. The display control unit 45 also causes the display unit 20 to display a setting screen and a progress status when analyzing the time-series data of the analysis target. However, the display contents on the display unit 20 are not particularly limited.

[0037] <2. Example of operation> So far, we have described an example of the configuration of data analysis apparatus 100. Next, we will describe an example of the operation of data analysis apparatus 100.

[0038] FIG. 3 is a flowchart showing an example of a data analysis method executed by processor 40 of data analysis apparatus 100. When the analysis process of the time series data is started, the data processing unit 41 selects an RNN to which the time series data to be analyzed is input from among a plurality of RNNs stored in the storage unit 30 (step S11). The selection of the RNN may be a selection of the type of RNN, or a selection of the number of layers of the RNN or the number of input data. The RNNs selectable by the data processing unit 41 may be predetermined according to the time series data to be analyzed. The RNNs may also be set according to the user's selection. The order of the RNNs to be set is not particularly limited.

[0039] Next, the data processing unit 41 acquires the time series data to be analyzed via the input unit 10, and inputs the data to the RNN set in step S11 to acquire time series prediction data (step S13). At this time, the data processing unit 41 sequentially inputs the input data, the number of which corresponds to the set RNN, to the RNN to acquire time series prediction data. The data processing unit 41 stores the acquired prediction data in the memory 47.

[0040] Next, the data processing unit 41 compares the acquired time series prediction data with the input time series data to be analyzed, and obtains the reproducibility of the time series data by the RNN (step S15). For example, the data processing unit 41 can obtain the reproducibility by the error function Err(X) shown in the above formula (2). The data processing unit 41 stores the obtained reproducibility information in the memory 47.

[0041] Next, the data processing unit 41 judges whether the determination of the reproducibility for all RNNs has been completed (step S17). If the determination of the reproducibility for some RNNs has not been completed (S17 / No), the data processing unit 41 returns to step S11, selects an RNN to which the time-series data to be analyzed is input from among the RNNs for which the determination of the reproducibility has not been completed, and executes the processes of steps S13 to S15. The data processing unit 41 repeatedly executes the processes of steps S11 to S17 until it is judged in step S17 that the determination of the reproducibility for all RNNs has been completed.

[0042] When the determination of the reproducibility for all RNNs is completed (S17 / Yes), the analysis unit 43 analyzes the characteristics of the time-series data to be analyzed based on the reproducibility information for each RNN stored in the memory 47 (step S19). The analysis unit 43 obtains information on the known input / output characteristics of each RNN stored in the storage unit 30, and analyzes the characteristics of the time-series data to be analyzed from the reproducibility for each RNN.

[0043] Next, the display control unit 45 generates an image signal for displaying the analysis result, outputs it to the display unit 20, and causes the display unit 20 to display the analysis result (step S21). The method for displaying the analysis result is not particularly limited. For example, a method of converting the characteristics of the time-series data into text and displaying it may be used, or a method of showing an analysis value for each characteristic item may be used.

[0044] <3.Effects> As described above, according to the data analysis device and data analysis method of the present embodiment, the characteristics of the time series data to be analyzed are analyzed based on the result of comparing the time series data to be analyzed with the time series data to be analyzed, which is obtained by inputting the time series data to be analyzed into a plurality of RNNs that have previously learned a plurality of teacher data related to the time series data to be analyzed, and the input / output characteristics of each RNN. In this embodiment, the reproducibility of the predicted data is determined as a result of the comparison, and the characteristics of the time series data to be analyzed are analyzed based on the reproducibility and the input / output characteristics of each RNN. This makes it possible to analyze the characteristics of the time series data based on the affinity with the input / output characteristics of the RNN. In particular, the strategic depth of the time series data to be analyzed according to the input / output characteristics of the RNN can be analyzed. Therefore, the analysis results of the time series data to be analyzed can be used for purposes such as analyzing the play of a video game or other game, recognizing and automatically editing noteworthy scenes such as highlight scenes, and summarizing the impact of noteworthy scenes on the whole.

[0045] According to the data analysis device and the data analysis method of the present embodiment, the multiple RNNs may include multiple RNNs of the same type that are different in at least one of the number of layers or the number of input data. By using such RNNs, it becomes possible to analyze the time series data to be analyzed even when the types of RNNs prepared in advance are limited.

[0046] Furthermore, according to the data analysis device and data analysis method of the present embodiment, the multiple RNNs are neural networks that have learned teacher data until time-series prediction data for a predetermined time-series input data becomes stable a predetermined number of times or for a predetermined period of time. This makes it possible to perform analysis based on the results of inputting the time-series data to be analyzed into RNNs whose input / output characteristics are stable, thereby increasing the reliability of the analysis results.

[0047] <4. Other embodiments> In the data analysis device 100 according to the embodiment described above, a plurality of RNNs that have learned teacher data in advance are stored in the storage unit 30, and the time-series data to be analyzed is input to these RNNs to determine the reproducibility of each RNN, but the data analysis device 100 according to the present invention is not limited to this example. For example, the time-series data to be analyzed may be used as learning data to learn an RNN, and the time-series data may be analyzed.

[0048] In the data analysis apparatus 100, an untrained RNN may be stored in the storage unit 30. Note that the stored RNN may be an RNN that has been trained a small number of times (has a low degree of completion) instead of an untrained RNN.

[0049] FIG. 4 is a flowchart showing an example of another data analysis method executed by processor 40 of data analysis apparatus 100. When the analysis process of the time series data is started, the data processing unit 41 selects an RNN to which the time series data to be analyzed is input from among the multiple RNNs stored in the storage unit 30, similar to step S11 above (step S31).

[0050] Next, the data processing unit 41 acquires the time-series data to be analyzed via the input unit 10, and repeatedly inputs the data to the RNN set in step S11 to execute a learning process (step S33). At this time, the data processing unit 41 sequentially inputs the input data to the RNN, the number of pieces of data corresponding to the set RNN.

[0051] Next, the data processing unit 41 judges whether the output time series prediction data has become stable for a predetermined number of times or for a predetermined time (step S35). In this embodiment, the data processing unit 41 compares the output time series prediction data with the analysis target time series data every time the analysis target time series data is input to the RNN, obtains the reproducibility of the time series data by the RNN, and judges whether the change in the reproducibility is within a predetermined range for a predetermined number of times or for a predetermined time. However, the method of judging whether the output time series prediction data is stable is not limited to this example.

[0052] If it is determined that the time series prediction data to be output is not stable (S35 / No), the data processing unit 41 returns to step S33 and repeats learning of the time series data to be analyzed. On the other hand, if it is determined that the time series prediction data to be output is stable (S35 / Yes), the data processing unit 41 stores information on the reproducibility of the time series data by the RNN in the memory 47 (step S37).

[0053] Next, the data processing unit 41 judges whether the determination of the reproducibility for all RNNs has been completed (step S39). If the determination of the reproducibility for some RNNs has not been completed (S39 / No), the data processing unit 41 returns to step S31, selects an RNN to which the time-series data to be analyzed is input from among the RNNs for which the determination of the reproducibility has not been completed, and executes the processes of steps S33 to S37. The data processing unit 41 repeatedly executes the processes of steps S31 to S39 until it is determined in step S39 that the determination of the reproducibility for all RNNs has been completed.

[0054] When the determination of the reproducibility for all RNNs is completed (S39 / Yes), the analysis unit 43 analyzes the characteristics of the time-series data to be analyzed based on the reproducibility information for each RNN stored in the memory 47 (step S41). The analysis unit 43 obtains information on the known input / output characteristics of each RNN stored in the storage unit 30, and analyzes the characteristics of the time-series data to be analyzed from the reproducibility for each RNN.

[0055] Next, the display control unit 45 generates an image signal for displaying the analysis result, outputs it to the display unit 20, and causes the display unit 20 to display the analysis result (step S43). The method for displaying the analysis result is not particularly limited. For example, a method of converting the characteristics of the time-series data into text and displaying it may be used, or a method of showing an analysis value for each characteristic item may be used.

[0056] As described above, the data analysis device 100 can also learn the RNN using the time-series data to be analyzed as learning data, and analyze the time-series data based on the reproducibility of the predicted data when the output predicted data becomes stable. By analyzing the time-series data in this way, the characteristics of the time-series data can be analyzed based on the affinity with the input / output characteristics of the RNN. In particular, since the RNN is trained using only the time-series data to be analyzed as the learning object, predictions can be made that are more specialized for the analysis object. Therefore, plays that are recognized as tricky for a particular player's play, such as plays that are thought to include tactics or special intentions, can be highlighted. Alternatively, plays that are thought to be simple mistakes or other normal plays can be distinguished from special plays.

[0057] Although the preferred embodiment of the present invention has been described in detail above with reference to the accompanying drawings, the present invention is not limited to such an example. It is clear that a person having ordinary knowledge in the technical field to which the present invention pertains can conceive of various modified or altered examples within the scope of the technical ideas described in the claims, and it is understood that these also naturally belong to the technical scope of the present invention.

[0058] For example, in the above embodiment, a plurality of RNNs and the input / output characteristics of each RNN are stored in advance in the storage unit 30 provided in the data analysis device 100, and the data processing unit 41 executes data processing by acquiring the RNNs from the storage unit 30, but the present invention is not limited to such an example. At least one piece of information regarding the plurality of RNNs and the input / output characteristics of each RNN may be stored in another device or storage medium connected to the data analysis device 100, and the data processing unit 41 may execute data processing by acquiring the information regarding the RNNs and the input / output characteristics of the RNNs from the other device or storage medium. [Explanation of symbols]

[0059] 10...input unit, 20...display unit, 30...storage unit, 40...arithmetic processing device, 41...data processing unit (acquisition unit), 43...analysis unit (acquisition unit), 45...display control unit, 47...memory, 100...data analysis device

Claims

1. In a data analysis device (100) for analyzing characteristics of time-series data, an acquisition unit (41, 43) that acquires time series data to be analyzed, time series prediction data obtained by inputting the time series data to be analyzed into a neural network used for time series prediction, and information regarding input / output characteristics of the neural network; a data processing unit (41) for comparing the time series prediction data with the time series data to be analyzed; an analysis unit (43) that analyzes characteristics of the time series data to be analyzed based on a result of comparing the time series prediction data for each of the plurality of neural networks with the time series data to be analyzed and information on input / output characteristics of each of the neural networks; A data analysis device comprising:

2. 2. The data analysis device according to claim 1, wherein the analysis unit (43) determines, for each of the neural networks, a degree of reproducibility of the time series prediction data for the time series data to be analyzed, and analyzes characteristics of the time series data to be analyzed based on the degree of reproducibility of the time series prediction data of each of the neural networks and information on input / output characteristics of each of the neural networks.

3. The data analysis apparatus according to claim 1 , wherein the plurality of neural networks include neural networks differing in at least one of the number of layers and the number of input data.

4. The data analysis device according to any one of claims 1 to 3, wherein each of the neural networks is a neural network that has learned training data until time series prediction data for a predetermined time series of input data becomes stable a predetermined number of times or for a predetermined period of time.

5. A data analysis method executed by a data analysis device (100) for analyzing characteristics of time-series data, comprising: acquiring time series data to be analyzed, time series prediction data obtained by inputting the time series data to be analyzed into a neural network used for time series prediction, and information on input / output characteristics of the neural network; comparing the time series forecast data with the time series data to be analyzed; analyzing characteristics of the time series data to be analyzed based on a result of comparing the time series prediction data for each of the plurality of neural networks with the time series data to be analyzed and information on input / output characteristics of each of the neural networks; , and a data analysis method.

Images