Information processing device, factor estimation method, training data generation method, and program

Abstract

This information processing device comprises: a first processing unit for acquiring a plurality of feature amounts that vary in time series from a communication system, and generating, for each combination of two feature amounts, a relationship feature amount representing the relationship between the feature amounts; a second processing unit for estimating the occurrence or the non-occurrence of a prescribed event on the basis of the relationship feature amount; and a third processing unit for estimating a factor of the prescribed event on the basis of one or more relationship feature amounts which serve as a basis for the estimation result of the occurrence of the prescribed event.

Description

Information Processing Apparatus, Factor Estimation Method, Learning Data Generation Method, and Program

[0001] The present invention relates to a technique for estimating a factor when the communication quality of an application deteriorates in a system that transmits application data via a communication network.

[0002] In recent years, the development of autonomous driving technology has advanced, and level 4 road driving in which the system assumes the driving entity is being realized under specific conditions. On the other hand, due to legal and safety reasons, remote monitoring by video is always required during autonomous driving, and it is necessary to transmit the video clearly and without interruption via an unstable mobile network.

[0003] In order to achieve stable video transmission, application-specific controls such as quality prediction of the line, line control using a plurality of communication networks in combination, and dynamic control of the bit rate and frame rate in the case of video are simultaneously and comprehensively implemented. [[ID=ll]]

[0004] When control is implemented in a plurality of layers, when the quality of the application layer deteriorates, it may be a complex factor caused by the control of the plurality of layers, and it is difficult to immediately identify the factor manually and change to appropriate control.

[0005] As a conventional technique for estimating a factor, Non-Patent Document 1 discloses a technique for calculating the importance of input features of machine learning and estimating a factor based on the importance.

[0006] In the conventional technique disclosed in Non-Patent Document 1, in a video streaming service, for the factor when viewing of a video is stopped based on feature amounts such as application quality and user behavior, it is determined whether it is due to the quality of the streaming video or due to the content (inconsistency with the user's preference).

[0007] Specifically, the causal relationship between feature amounts is analyzed manually, and by performing analysis using Asymmetric Shapley Value that can calculate the importance of feature amounts in consideration of the causal relationship, the feature amounts (factors) that have affected the classification of machine learning are analyzed.

[0008] Sosa Akimoto, et al. Quantitative causality analysis of viewing abandonment reasons using Shapley valueKumar, et al. Image data augmentation approaches: A comprehensive survey and future directions.Dwibedi, et al. Cut, Paste and Learn: Surprisingly Easy Synthesis for Instance Detection

[0009] As described above, conventional techniques manually analyze causal relationships between features, but they assume that these relationships are fixed across all features and do not change over time. Furthermore, conventional techniques cannot consider the time-series nature of multiple logs, making it impossible to analyze the factors involved when classifying KPIs (Key Performance Indicators, such as quality-related or content-related video interruptions) by capturing the temporal changes in input features that occurred in the past.

[0010] Therefore, when conventional technologies are applied to use cases such as the degradation of application quality in remote control of autonomous driving, it is expected that the contributing factors will not be able to be analyzed effectively.

[0011] This invention has been made in view of the above points, and aims to provide a technology that enables factor analysis of a predetermined event, taking into account changes in the relationships between feature quantities over time.

[0012] According to the disclosed technology, an information processing device is provided, comprising: a first processing unit that acquires multiple feature quantities that change over time from a communication system and generates relational feature quantities that express the relationship between the feature quantities for each combination of two feature quantities; a second processing unit that estimates whether or not a predetermined event has occurred based on the relational feature quantities; and a third processing unit that estimates the factors of the predetermined event based on one or more relational feature quantities that formed the basis of the estimation result of whether or not the predetermined event has occurred.

[0013] The disclosed technology provides a method that enables factor analysis of a given event, taking into account changes in the relationships between features over time.

[0014] This is a diagram illustrating the definition of factors. This is a diagram illustrating the problems in the prior art. This is a diagram illustrating the image of factor analysis in the prior art. This is a diagram illustrating the image of factor analysis in an embodiment of the present invention. This is a diagram illustrating an example of the functional configuration of the information processing device 100 in the first embodiment. This is a diagram illustrating the image of the processing of the pre-processing unit 110. This is a flowchart illustrating the operation of the pre-processing unit 110. This is a diagram illustrating the image of the processing performed by the pre-processing unit 110. This is a diagram illustrating the image of the processing of the importance calculation unit 130. This is a flowchart illustrating the operation of the post-processing unit 140. This is a diagram illustrating a specific example of the processing of the post-processing unit 140. This is a diagram illustrating an example of the functional configuration of the information processing device 100 in the second embodiment. This is a diagram illustrating an example of input and output to the data expansion unit 150. This is a processing flow related to the determination of the number of transfers. This is a table showing comparative examples of accuracy and calculation time depending on the presence or absence of data expansion and the difference in the number of applications. This is a diagram illustrating an example of the functional configuration of the information processing device 200. This is a diagram illustrating an example of the hardware configuration of the device.

[0015] Hereinafter, embodiments of the present invention (this embodiment) will be described with reference to the drawings. The embodiments described below are merely examples, and the embodiments to which the present invention is applied are not limited to the embodiments described below.

[0016] In the following embodiments, the technology for analyzing the causes of video interruptions in remote monitoring will be described, but video interruptions are merely one example. The technology according to the present invention is applicable to a variety of phenomena, not limited to video interruptions.

[0017] In the following, we will first explain the definition of the factors in this embodiment, then describe the problems in more detail, and then describe the technology related to this embodiment in detail.

[0018] (Definition of Factors) First, referring to Figure 1, we will explain the definition of factors for the interruption (stoppage) of remote monitoring video. As shown in Figure 1, a factor is the earliest event in a causal relationship. Video interruption occurs through a chain of factors and influencing events. Note that "events" also include "settings".

[0019] In other words, a contributing factor is the initial event that caused the video to freeze, and resolving or avoiding it will reduce the likelihood of the video freezing to the greatest extent possible.

[0020] The influencing events shown in Figure 1 are events that have a causal relationship with a factor (events caused by the factor). Examples of influencing events include increased latency, increased packet loss rate, and decreased bitrate.

[0021] (Regarding the problem) As mentioned above, in the prior art disclosed in Non-Patent Document 1, the causal relationships between features are manually analyzed and defined in the stage before constructing the machine learning model. The importance of the features (factors in classification and prediction results) is calculated based on the assumption that the relationships in these causal relationships are fixed across all features, regardless of the passage of time.

[0022] In addition, the prior art disclosed in Non-Patent Document 1 does not take into account the time-series nature of multiple logs, so when classifying KPIs (for example, quality-related / content-related determination tasks or video interruptions), it is not possible to capture changes in input features that have occurred in the past and use them as factors.

[0023] On the other hand, in the case of application quality (video quality) degradation in autonomous driving remote control systems, the relationships between input features may change over time. For example, at certain times, increased latency may be caused by deterioration of wireless quality, leading to video interruptions, while at other times, increased latency may be caused by inappropriate settings of the video bitrate, leading to video interruptions.

[0024] Furthermore, when implementing control using multiple layers, value fluctuations chain together among multiple logs (=input features), meaning that changes in input features may have occurred in the past before the KPI classification was performed. For example, if latency increases and video interruption occurs after a decrease in wireless quality, the decrease in wireless quality → video interruption will occur within a few milliseconds to a few seconds. If the system cannot tolerate a delay of a few seconds in events, it may not be able to identify factors such as "decreased wireless quality."

[0025] However, as mentioned above, the prior art disclosed in Non-Patent Document 1 does not take into account the time-series nature of multiple logs, and therefore cannot capture and use past changes in input features as a factor when classifying KPIs. Consequently, it is expected that if this prior art is applied to use cases such as the deterioration of application quality in remote control of autonomous driving, it will not be possible to effectively analyze the factors.

[0026] Refer to Figure 2 to explain the challenges in conventional technology in more detail. Figure 2 shows examples of the relationship between features (events) and KPIs (relationships focusing on time), including "simultaneous," "time-direction shift," "chained," "composite," and "chained + composite."

[0027] "Simultaneous" indicates that feature 3 and KPI occurred at the same time. "Time-shifted" indicates that KPI occurred after a certain amount of time had passed since feature 3 occurred. "Chain" indicates that there is a chain reaction: feature 1 -> feature 2 -> feature 3 -> KPI. "Combined" indicates that both feature 1 and feature 3 are factors in KPI. "Chain + Combined" indicates that both chaining and combined reactions are occurring.

[0028] Conventional technology cannot account for relationships that change over time. Therefore, while conventional technology can handle "simultaneous" events, it struggles to handle other types of relationships. However, for "time-direction shifts," conventional technology may be able to handle them if the shift magnitude is small. For this reason, it is indicated with a "△" in Figure 2.

[0029] (Outline of the Embodiment) In order to solve the above problems, in this embodiment, the information processing device 100, which will be described later, performs preprocessing, prediction of video interruption, calculation of the degree of contribution (importance) to the video interruption, and postprocessing. Of these processes, for the prediction of video interruption and the calculation of the degree of contribution to the video interruption, existing technologies disclosed in Non-Patent Document 1, for example, can be used.

[0030] In preprocessing, the information processing device 100 analyzes the relationships between input features (logs obtained from the system) for the machine learning model used to predict video interruptions in a time series, and expresses the results as vectors. In postprocessing, the information processing device 100 estimates the factors based on the output of the machine learning model and the importance of the vectors generated in preprocessing.

[0031] For comparison, Figure 3 shows an image of the factor analysis of the conventional technology. Figure 4 shows an image of the factor analysis in this embodiment.

[0032] As illustrated in Figure 4, in the preprocessing stage, the information processing device 100 obtains subvectors from the feature quantities in time units, determines whether there is a relationship between each pair of feature quantities, and determines the value of the generated vector according to the determination result. This process is repeated while sliding the acquisition position of the subvectors in the time direction.

[0033] In post-processing, the information processing device 100 performs a threshold-based determination of the importance of the vectors generated in pre-processing, and analyzes the factors from the content of the vectors that remain after the threshold determination.

[0034] The first and second embodiments will be described below.

[0035] [First Embodiment] The configuration and operation of the information processing device 100 will be described in detail below.

[0036] (Example of Functional Configuration of Information Processing Device 100) Figure 5 shows an example of the functional configuration of the information processing device 100 in this embodiment. As shown in Figure 5, the information processing device 100 has a pre-processing unit 110, an estimation unit 120, an importance calculation unit 130, and a post-processing unit 140.

[0037] The pre-processing unit 110 may also be called the first processing unit. The estimation unit 120 may also be called the second processing unit. The "importance calculation unit 130 + post-processing unit 140" may also be called the third processing unit. The general operation of each unit is as follows.

[0038] The preprocessing unit 110 performs causal relationship inference in a time series as a preprocessing step. Specifically, the preprocessing unit 110 performs the processing described later on the logs (features) acquired from the autonomous driving remote control system and generates a vector that represents the relationships between the features. The autonomous driving remote control system is an example of a communication system.

[0039] The prediction unit 120 performs prediction of video interruptions. Specifically, the prediction unit 120 is a time-series machine learning model that learns a vector formed by combining the vector generated in preprocessing with KPI (video interruption information), using vectors other than KPI as explanatory variables and the KPI vector as the target variable. The prediction unit 120 classifies or predicts video interruptions as binary (interruption present / interruption absent) and outputs the result. In this embodiment, a neural network such as RNN, LSTM, or LSTM + attention is used as the machine learning model that handles time series.

[0040] The importance calculation unit 130 calculates the importance of the features related to the video interruption. Specifically, when the estimation unit 120 outputs a binary value, the importance calculation unit 130 calculates which features were given more weight (which features contributed to the judgment) from among the input features used as the basis for that value. As a method for calculating feature importance, for example, Shapley Value or LIME can be used.

[0041] The post-processing unit 140 performs a causal relationship analysis as a post-processing step. Specifically, the post-processing unit 140 uses the classification and prediction results of video interruptions and the output of the feature importance calculation at the time of video interruption to determine the features that are thought to have influenced the video interruption through the processing described later, and estimates the factors.

[0042] The operation of each part will be explained in detail below.

[0043] (Preprocessing Unit 110) First, the preprocessing unit 110 will be described. The preprocessing unit 110 performs the following processing on the log (referred to as a feature amount) acquired from the automatic driving remote control system, and generates a vector representing the relationship between the feature amounts. FIG. 6 shows an image of the processing of the preprocessing unit 110.

[0044] First, the preprocessing unit 110 acquires partial vectors (the shaded frame portion in FIG. 6) with a specific time width for two feature amounts, and determines the presence or absence of a causal relationship for the two acquired partial vectors.

[0045] The method for determining the presence or absence of a causal relationship (also referred to as causality) is not limited to a specific method. For example, existing techniques such as vector autoregression (VAR) and Granger test can be used.

[0046] When the preprocessing unit 110 determines that there is a causal relationship between two feature amounts, it calculates the harmonic mean of the values of each vector and sets it as the value of the generated vector. Note that using the harmonic mean is just an example, and other means (e.g., arithmetic mean, weighted mean) may also be used.

[0047] When the preprocessing unit 110 determines that there is no causal relationship between two feature amounts, it sets a pre-set specific value (such as -1) as the value of the generated vector.

[0048] The name of the vector newly generated by the preprocessing unit 110 needs to be able to distinguish the names of the two source vectors. Therefore, in this embodiment, as shown in FIG. 6, the vector generated from feature amount 1 and feature amount 2 is denoted as (1, 2). The same applies to other combinations of two feature amounts. Note that KPI is also one of the feature amounts.

[0049] The preprocessing unit 110 executes the above processing for all combinations of feature amounts.

[0050] (Processing Flow of Preprocessing Unit 110) The operation of the preprocessing unit 110 will be described more specifically with reference to the flowchart of FIG. 7. In the following description, it is assumed that there are feature amount 1, feature amount 2, and KPI as feature amounts obtained as logs.

[0051] <S101 (Step 101)> The preprocessing unit 110 acquires and inputs the time-series data for each feature. The preprocessing unit 110 also inputs the size (time width) of the subvectors.

[0052] For example, the preprocessing unit 110 obtains "a1, a2, ..., at, ..., aT" as time series data (time t=1 to T) for feature quantity 1 (e.g., wireless quality), "b1, b2, ..., bt, ..., bT" as time series data for feature quantity 2 (e.g., inflight), and "k1, k2, ..., kt, ..., kT" as time series data for KPI (e.g., video interruption). For example, kt=0 represents no video interruption, and kt=1 represents video interruption.

[0053] <S102> The preprocessing unit 110 obtains a subvector of the above size for each feature quantity.

[0054] For example, if the time width is assumed to be 3, the preprocessing unit 110 obtains (a1, a2, a3) as a subvector of feature quantity 1, (b1, b2, b3) as a subvector of feature quantity 2, and (k1, k2, k3) as a subvector of KPI.

[0055] <S103, S104> The preprocessing unit 110 determines whether or not there is a causal relationship for each combination of two subvectors, for example, by using vector autoregression (VAR) and Granger tests.

[0056] Specifically, the causal relationships between (a1, a2, a3) and (b1, b2, b3), between (b1, b2, b3) and (k1, k2, k3), and between (k1, k2, k3) and (a1, a2, a3) are determined.

[0057] For combinations where a causal relationship is determined, the process proceeds to S106; for combinations where a causal relationship is determined, the process proceeds to S105.

[0058] <S105, S106> In S105, when it is determined that there is no causal relationship, the preprocessing unit 110 sets, for example, (-1, -1, -1) as the value of the vector to be generated for that combination.

[0059] In S106, if a causal relationship is determined to exist, the preprocessing unit 110 sets the harmonic mean of the vectors of the combination as the value of the vector to be generated.

[0060] For example, if it is determined that there is a causal relationship between (a1, a2, a3) and (b1, b2, b3), the preprocessing unit 110 sets the vectors (harmonic mean of a1 and b1, harmonic mean of a2 and b2, harmonic mean of a3 and b3) as the values ​​of the vectors to be generated.

[0061] <S107> The preprocessing unit 110 performs the processes S102 to S106 with a time interval shift. That is, if the time interval is 3, the processes 102 to S106 are performed for each of the time intervals of time = 2, 3, 4; time = 3, 4, 5; ...; time = T-2, T-1, T.

[0062] The preprocessing unit 110 may also perform normalization on the vector obtained by the above process.

[0063] Figure 8 shows an image of the processing performed by the preprocessor 110 for the aforementioned "simultaneous," "time-direction shift," "chaining," and "combination" processes.

[0064] Hereafter, the vector that shows the relationship between two subvectors generated during preprocessing will be called a "relational feature."

[0065] (About the prediction unit 120) The prediction unit 120 is a time-series machine learning model that learns using a vector formed by combining the vectors (relational features) generated in preprocessing with KPI (video interruption information), with the vectors other than KPI (relational features) as explanatory variables and the KPI vector as the target variable. The trained machine learning model takes the relational features as input, classifies or predicts video interruptions as binary (interruption present / interruption absent), and outputs the result.

[0066] In this embodiment, it is assumed that RNN or LSTM, etc., are used as the time-series machine learning model. Furthermore, it is assumed that this machine learning model is a supervised learning model (training data: KPI vectors).

[0067] The inference unit 120 may either hold a machine learning model or be a machine learning model itself.

[0068] The following describes the processing in the learning phase and the inference phase of a machine learning model. In the learning phase, the inference unit 120 is assumed to have a learning function and hold the machine learning model to be learned. The inference unit 120 in the learning phase may also be called the "learning unit". Here, too, the explanation assumes that there is feature 1, feature 2, and KPI.

[0069] <Learning Phase> The three relational features generated in preprocessing for each time interval are denoted as (1,2), (1,KPI), and (2,KPI). Since (1,2), (1,KPI), and (2,KPI) are generated for each time interval, the relational features for time interval k are denoted as (1,2)k, (1,KPI)k, and (2,KPI)k. For example, for the first time interval 1, the features are (1,2)1, (1,KPI)1, and (2,KPI)1.

[0070] In the learning phase, the inference unit 120 first inputs a relational feature with a time interval of 1 (for example, (1, KPI)1) into the machine learning model. The machine learning model outputs a prediction result (classification result) of whether or not there was a video interruption within a time interval of 1. The inference unit 120 updates the parameters of the machine learning model so that the error between this output and the correct answer (whether or not there was actually a video interruption within a time interval of 1) is reduced.

[0071] Next, a relational feature with a time interval of 2 (e.g., (1, KPI)2)) is input to the machine learning model, and the parameters of the machine learning model are updated to minimize the error between the output and the correct answer (whether or not there was actually a video interruption with a time interval of 2). This process is performed for data with a time interval of 2. For each time interval k, one relational feature is used from among the three relational features "(1, 2)k, (1, KPI)k, (2, KPI)k" that is appropriate for the characteristics of the data.

[0072] <Inference Phase> In the inference phase, for example, when predicting whether or not there was a video interruption based on (1, KPI)k, (1, KPI)k is input to the trained machine learning model, and the machine learning model outputs a prediction result indicating whether or not there was a video interruption within a time interval k.

[0073] (Regarding the importance calculation unit 130) When the inference unit 120 outputs a binary result which is the prediction result of whether or not there was a video interruption, the importance calculation unit 130 calculates which of the input relational features were given more weight (which relational features contributed to the judgment) from among the relational features that were used as the basis for that result.

[0074] Representative methods for calculating such importance include Shapley Value and LIME. These methods provide explainability to many deep learning models, including RNNs and LSTMs, and allow the relational features that were most relevant in deriving the prediction / classification results to be expressed by their numerical magnitude.

[0075] Shapley Value or LIME approximates a Black-Box model to the feature vector (relational feature) being analyzed using a White-Box model that can explain the neighborhood of the explanatory variable. It then compares the classification / prediction results for each combination of features (for example, (1,2), (2,3), and (1,3) in the case of feature 1, feature 2, and feature 3) and calculates importance based on which relational feature caused the most change in the classification / prediction results.

[0076] Figure 9 shows an image illustrating how the importance calculation unit 130 calculates the importance of each related feature based on the classification / prediction results for video interruption. Note that #1, #2, and #3 in the lower left of Figure 9 indicate the order of importance.

[0077] (Regarding the post-processing unit 140) The post-processing unit 140 determines the relative features that are thought to have influenced the video interruption based on the classification and prediction results of the video interruption and the importance of the relative features at the time of the video interruption, and estimates the factors by processing as described later. The processing of the post-processing unit 140 will be explained according to the procedure in the flowchart of Figure 10. In addition, the naming convention for relative features (e.g., (1,2) is generated from feature 1 and feature 2) follows the one explained above.

[0078] <S201> First, the post-processing unit 140 calculates the average importance value as a threshold for each related feature in the negative examples (in this embodiment, related features that did not cause video interruption are considered negative examples). The threshold is calculated for each related feature.

[0079] <S202> In this embodiment, relational features determined to be video interruptions by the machine learning model are treated as positive examples. For the positive examples, the post-processing unit 140 uses the importance calculation unit 130 to calculate the importance of the relational features. As mentioned above, it is assumed that the importance of the relational features will be calculated using Shapley or LIME.

[0080] <S203> The post-processing unit 130 extracts only the relational features whose importance is "positive and above the threshold" from among the multiple relational features whose importance was calculated in S202. Note that if the threshold is positive, "positive and above the threshold" may be written as "above the threshold". The threshold here is the threshold calculated in S201. Steps S204 and S205 below are performed on the relational features extracted in S203. In S204 and S205, factors are obtained by combining the relational features remaining after the threshold determination with features from the same generator. Specifically, it is as follows.

[0081] <S204> The post-processing unit 140 checks for matching log names (feature 1, feature 2, etc.) that constitute the relational features and constructs a tree structure.

[0082] <S205> The post-processing unit 140 determines the log name (leaf in the tree structure) furthest from the video interruption as the cause and outputs that cause.

[0083] (Specific examples of S204 and S205) Specific examples of S204 and S205 will be explained with reference to Figure 11. For example, suppose that in time interval 3 (the third time interval), the prediction result is "video interruption occurred," and that "(2, KPI) and (1, 2)" in time interval 3 are extracted as relational features whose importance is above a threshold. Figure 11 shows that (2, KPI) and (1, 2) have been extracted.

[0084] At this point, since there is a feature named "Feature 2" from the same source, it can be combined to construct a tree structure of "KPI -> Feature 2 -> Feature 1". Alternatively, as shown in Figure 11, this tree structure can be written as "Feature 1 -> Feature 2 -> KPI". The post-processing unit 140 determines that Feature 1, which is furthest from the video interruption, is the cause. This process is performed for each time interval in which the prediction result is "video interruption present".

[0085] Furthermore, if, for example, (1, KPI) and (2, KPI) are extracted in S203, then feature 1 and feature 2 are each considered as factors.

[0086] Furthermore, for example, if there are features 1, 2, and 3 other than KPI, and the relational features (1, KPI), (1, 2), and (1, 3) are extracted as having an importance above a threshold, then the following tree structure is obtained (intended to have two branches from feature 1 to features 2 and 3).

[0087] KPI -> Feature 1 -> Feature 2 -> Feature 3 In this case, the post-processing unit 140 determines that each of Feature 2 and Feature 3 is a factor.

[0088] [Second Embodiment] Next, a second embodiment will be described. In the following, the prior art and problems of the second embodiment will be described, and then the method of the second embodiment will be described in detail.

[0089] (Prior Technology) In fields such as image recognition and natural language processing, data augmentation has been proposed to resolve imbalances in training data in supervised learning and to improve classification and prediction performance. Specifically, as a data augmentation targeting images, a technique is known that changes the appearance of input data by rotating, flipping, scaling, cropping, color conversion, etc., and generates diverse training samples from a small amount of training data (for example, Non-Patent Document 2).

[0090] (Challenges of Conventional Technology) However, when targeting operational logs in real-world fields, such as remote monitoring of autonomous driving, the data format differs from that of images, etc., making it impossible to directly apply the conventional data augmentation methods described above.

[0091] Furthermore, quality degradation events such as video interruptions in real-world field use are rare, resulting in small amounts of interruption data that tend to be unbalanced and heavily skewed compared to non-interruption data. In addition, assigning interruption labels involves matching multiple layers of logs—video, communication, terminals, applications, etc.—to determine the interruption interval, making labeling difficult. In short, it is difficult to stably collect a sufficient amount of reliable interruption-labeled data from real-world operations, which is a prerequisite for supervised learning.

[0092] As a result, the interruption detection model suffers from insufficient training of the interruption class, making it prone to overlooking interruptions and resulting in low recall rates for interruption detection. Furthermore, in quality degradation factor estimation, the small number of interruption events leads to instability in the estimation of training results and feature importance (e.g., Shapley), resulting in a large amount of noise (inclusion of incorrect factors) in the factor estimation results, making it difficult to use in practical applications.

[0093] In other words, estimating quality degradation factors in real-world applications presents challenges in both areas: (1) securing training data to improve the recall rate of interruption detection, and (2) developing training data to stabilize the factor estimation results at a quality level suitable for operational use.

[0094] The technology according to the second embodiment is proposed in view of the above-mentioned problems, and aims to improve the accuracy of interruption detection and cause estimation when using supervised learning by augmenting training data related to interruptions, which rarely occur in the field, using data augmentation. In other words, it aims to suppress the decrease in recall rate and the instability of cause estimation (increase in noise) caused by the small amount and imbalance of interruption data, and to obtain stable estimation results that are useful for operation.

[0095] (Outline of the technology according to the second embodiment) In the second embodiment, to address the problem that interruption data is small and unbalanced due to video interruptions that rarely occur in the field, the training data used for supervised learning is augmented using data augmentation to improve the accuracy of interruption detection and cause estimation. In other words, the aim is to mitigate the problems of low recall in interruption detection and unstable (noisy) cause estimation results due to a lack of labeled interruption data by improving the training data.

[0096] Specifically, in the second embodiment, the concept of Copy & Paste (Non-Patent Literature 3), known as a type of data augmentation method, is applied to real-world time-series logs (multivariate time-series data consisting of multiple features). Copy & Paste is a method that artificially increases the number of training samples for a small class by copying (Copy) a data fragment containing an event belonging to a small class (in this embodiment, video interruption) and pasting (Paste) it onto other data (typically non-interruption data, which is a large class).

[0097] In this embodiment, Copy&Paste can handle interruptions not as "points" in time-series data, but as units of a predetermined time width (interval, window). That is, a continuous interval in which an interruption occurs (interrupted interval) is extracted as the source for copying, a predetermined interval in the non-interrupted data is selected as the destination for pasting, and at least a portion of the destination interval is transferred (replaced or combined) with data fragments based on the interrupted interval of the source for copying, thereby generating a new training sample that includes interruptions.

[0098] Furthermore, the synthesized samples generated by Copy & Paste are assigned labels corresponding to intervals containing interruptions (interruption labels). In other words, the synthesized samples are treated as "samples containing interruptions" in supervised learning, increasing learning opportunities for interruption classes. This helps to suppress overlooking interruptions and improve recall, while also stabilizing the calculation of importance used in factor estimation and reducing variability and noise in the estimation results. The device configuration and operation in the second embodiment will be described in detail below.

[0099] (Example of device configuration and operation overview in the second embodiment) Figure 12 is a diagram showing an example of the functional configuration of the information processing device 100 in the second embodiment. As shown in Figure 12, the information processing device 100 in the second embodiment has a configuration in which a data expansion unit 150 is added to the configuration of the information processing device 100 of the first embodiment shown in Figure 5. That is, as shown in Figure 12, the information processing device 100 of the second embodiment has a data expansion unit 150 in addition to a pre-processing unit 110, an estimation unit 120, an importance calculation unit 130, and a post-processing unit 140.

[0100] The data augmentation unit 150 receives logs and measured values ​​(explanatory variables) and target variables (e.g., presence or absence of video interruption) acquired from the system to be targeted for factor estimation (e.g., an autonomous driving remote monitoring system), applies data augmentation processing to the data to generate augmented explanatory variables and augmented target variables, and outputs them to the preprocessing unit 110, etc. In other words, the second embodiment is characterized in that, while basically maintaining the processing content from the preprocessing unit 110 onward (causality inference, time series machine learning, importance calculation, factor estimation by post-processing), the data augmentation unit 150 reinforces the learning data (explanatory variables and target variables) that serve as input.

[0101] Specifically, the data augmentation unit 150 improves class imbalance by transferring interrupted intervals, which represent a small number of classes in the input data, to uninterrupted intervals, which represent a large number of classes. Here, "transfer" is a concept that includes the operation of copying a predetermined time interval (interrupted interval) containing a small number of classes (interrupted) and pasting (replacing or inserting) it into another time interval (usually an uninterrupted interval). In other words, by artificially increasing the number (or proportion) of samples of small classes, the bias in the class distribution in supervised learning is mitigated.

[0102] <Specific Examples of Transfer> Table 1 in Figure 13 shows an example of input to the data expansion unit 150. Table 1 shows time-series data including multiple explanatory variables (features) such as "Company A's measured value," "Company A's predicted value," "Company B's measured value," and "Srtt," and the target variable, "video interruption" (e.g., 0: none, 1: present). Note that "Company A's measured value" refers to the measured value when using Company A's (mobile operator's) network. The same applies to the others.

[0103] Table 1 shows that there are few interrupted data points where the dependent variable is 1, and many non-interrupted data points where it is 0, meaning that the amount of interrupted data is small and unbalanced.

[0104] Table 2 in Figure 13 shows an example of the expanded data output by the data expansion unit 150. In Table 2, the transfer of interruption intervals included in Table 1 increases the amount of data containing interruption (1), and the class imbalance is mitigated. In other words, it becomes possible to increase the learning opportunities for interruption classes as data to be used for supervised learning.

[0105] If there are two or more consecutive interruptions, the data expansion unit 150 may treat the consecutive portion as a single interval (block) and transfer it in block units. In other words, instead of decomposing consecutive interruptions into scattered single events, the time-continuous pattern of interruptions (including the transition of measured values ​​before and after) can be reflected in the training data. This makes it easier to reflect the time-series characteristics of interruptions in the training.

[0106] (Detailed operation of the data extension unit 150) The interrupted data extension (copy and paste type transfer) process performed by the data extension unit 150 will be described in more detail below.

[0107] First, the data augmentation unit 150 receives logs and measurements (time-series data of multiple explanatory variables) and the target variable (video interruption information) acquired from the system targeted for factor estimation. Next, the data augmentation unit 150 extracts the intervals in which the target variable is interrupted (e.g., 1). If the interruptions are consecutive, the consecutive intervals are extracted as a single block.

[0108] The data extension unit 150 selects a destination section (paste destination) from a number of classes (non-interrupted, e.g., 0). The destination section may be selected randomly, or it may be selected based on conditions such as time zone, type of road the autonomous vehicle is traveling on, and communication conditions.

[0109] Next, the data augmentation unit 150 transfers the extracted interrupted intervals (source data) to the selected destination intervals (destination data). Specifically, it replaces the explanatory and dependent variables corresponding to the destination intervals with the explanatory and dependent variables corresponding to the interrupted intervals in the source data. Alternatively, the explanatory and dependent variables corresponding to the interrupted intervals in the source data may be inserted into the time-series data at the destination. In other words, by replacing some of the non-interrupted data with interrupted data, or by adding interrupted intervals, the number of interrupted samples (or the proportion of interrupted data) in the augmented data is increased.

[0110] The data augmentation unit 150 finalizes the transferred explanatory variable column and target variable column as augmented data and outputs them to the preprocessing unit 110. The preprocessing unit 110 generates relational features based on the augmented data, and the inference unit 120 (machine learning model) is trained using these relational features.

[0111] As described above, the data extension unit 150 may increase the sequence length by inserting source intervals without replacing the data at the pasting destination. In other words, both the "method of replacing non-disconnected data with disconnected data" and the "method of increasing the amount of data by adding disconnected intervals" are suitable for the purpose of reinforcing disconnected data.

[0112] (Regarding the number of transfers) In this embodiment, the transfer process described above is performed multiple times. Examples of the number of transfers and the amount of transfer are described below.

[0113] <Definition of Transfer Count> The transfer count is the number of times the data expansion unit 150 performs the operation of transferring the interrupted section from the copy source to the paste destination. In other words, one transfer corresponds to a series of operations in which at least one interrupted section (single or block) is extracted, one paste destination section is selected, and pasted (replaced or inserted) to the paste destination.

[0114] <Definition of Transfer Amount> The transfer amount (e.g., "25%") is a quantity that expresses the number of transfers in relative terms, and there may be multiple definitions of how the denominator is taken. In this embodiment, it may be defined as one (or more) of the following.

[0115] Definition Example 1: Example using the number of multi-class intervals as the denominator. In Definition Example 1, the transfer amount α[%] is defined as the "total number of data points in multi-class intervals that can be selected as pasting destinations" as the denominator. That is, for example, a transfer amount of 25% means that a transfer (replacement or addition) will be performed on 25% of the data points in the multi-class intervals that can be pasted.

[0116] Definition Example 2: Example using the total number of samples (total number of rows) as the denominator. In Definition Example 2, the transfer amount α[%] is defined with the total number of samples to be expanded (e.g., the total number of rows in Table 1) as the denominator. That is, for example, a transfer amount of 25% means that a number of data equivalent to 25% of the total number of data points will be replaced or added by transferring the discontinued intervals.

[0117] (Method for determining the number of transfers (transfer amount)) The number of transfers (or transfer amount) may be specified by the system user, or it may be determined by the data expansion unit 150 through exploration based on feedback from accuracy evaluation. The processing flow related to determining the number of transfers will be explained with reference to Figure 14.

[0118] <S301, S302> In S301, the data expansion unit 150 determines whether the user of the information processing device 100 has specified the number of transfers (or the amount of transfer). If the user has specified, the data expansion unit 150 repeats the transfer process according to the specified number of transfers (or the specified amount of transfer) to finalize the expanded data (S302).

[0119] <S303, S304> If no specification is made by the user, the data augmentation unit 150 sets multiple candidate transfer amounts (0%, 25%, 50%, etc. in the example of Figure 14) and searches for and determines the transfer amount based on the accuracy evaluation feedback from the estimation unit 120. Specifically, the augmented data generated by changing the transfer amount is used to train and evaluate the estimation unit 120, etc., and the search range is updated based on the results to determine a better transfer amount. In the example of Figure 14, binary search is used as the search method, but it is not limited to this.

[0120] <Examples of evaluation metrics in the search> In the search method described above, the data augmentation unit 150 determines the transfer amount that yields the best (maximum or minimum) predetermined evaluation metric based on the learning and evaluation results obtained for each candidate transfer amount. The evaluation metric may include at least one of the following.

[0121] (1) Recall rate for outage detection: When prioritizing the reduction of overlooked outages, maximize the recall rate as the primary indicator.

[0122] (2) F1 score: When prioritizing the balance between precision and recall, maximize the F1 as the primary indicator.

[0123] (3) Low noise in factor estimation: From the perspective of suppressing the inclusion of false factors, the variability (variance) of factor estimation results and the frequency of false extraction based on the difference from the mean of negative examples are used as indicators to minimize the noise.

[0124] In other words, by selecting the evaluation indicator, it is possible to switch between prioritizing "improving the performance of interruption detection" and "stabilizing factor estimation."

[0125] The data expansion unit 150 may use the above evaluation indicators individually, or it may use an integrated indicator that combines multiple indicators by weighting them (e.g., w1 × Recall + w2 × F1 + w3 × Noise Suppression, etc.). In other words, the optimization criteria for the transfer amount can be flexibly set according to operational requirements (minimizing oversights, suppressing false alarms, explaining the cause estimation, etc.).

[0126] In the example above, we described an example of optimizing the amount of transfer through exploration, but it is also possible to optimize the transfer location (insertion location / replacement location) in addition to the amount of transfer. That is, evaluation metrics are calculated for various transfer locations, and the transfer location that yields the best evaluation metrics is determined.

[0127] (Summary of the second embodiment) In the second embodiment, the information processing device 100 includes a data expansion unit 150 in addition to the preprocessing unit 110, estimation unit 120, importance calculation unit 130, and postprocessing unit 140. The data expansion unit 150 generates training data for supervised learning by performing a copy-and-paste type data operation that transfers interrupted sections extracted from the field log to other sections. In other words, it mitigates the "small amount and imbalance of interrupted data" caused by video interruptions that rarely occur in the field without significantly increasing system operation or labeling work.

[0128] Furthermore, in the second embodiment, the data augmentation unit 150 can determine the number of transfers (and insertion positions) in a way that improves the accuracy of the estimation unit 120 (machine learning model). Specifically, the data augmentation unit 150 may determine the number of transfers by user specification, or it may automatically optimize the amount of transfer by searching for it (e.g., binary search) using evaluation results (accuracy indicators) from the estimation unit 120, etc. In other words, the data augmentation conditions can be flexibly optimized according to operational requirements (minimizing oversights, suppressing false alarms, stabilizing factor estimation, etc.).

[0129] (Effects of the technology according to the second embodiment) According to the second embodiment, the data expansion unit 150 transfers interrupted intervals extracted from the field log to other intervals to generate training data, thereby ensuring sufficient balanced data for training without increasing field operation. In other words, even in situations where interruptions are rare and labeling is difficult, it becomes easier to secure the amount of interrupted data necessary for supervised learning.

[0130] Furthermore, according to the second embodiment, data augmentation can improve evaluation metrics (e.g., AUP, F1, etc.) in interruption detection / classification, and stabilize the output of factor estimation results. In other words, it suppresses problems such as low recall in interruption detection due to small amounts or imbalances, or high noise in factor estimation results, making it easier to obtain estimation results that are easy to handle in operation.

[0131] Figure 15 is a table showing comparative examples of accuracy and calculation time based on the presence or absence of data augmentation and the number of applications. The table shows Precision, Recall, and F1 for the conditions of "no interruption" and "with interruption," and also shows the overall classification performance index MCC (Matthews Correlation Coefficient) and factor calculation time (data augmentation time + estimation time).

[0132] In the example in Figure 15, (a) LSTM + Shap (a method that uses Shap for factor estimation), (b) a configuration without data augmentation (indicated as "Data augmentation ×" in the table), and (c) a configuration in which data augmentation is applied a certain number of times (indicated as "Applied 25 times each", "Applied 50 times each", and "Applied 75 times each" in the table).

[0133] As shown in Figure 15, when data augmentation is applied, accuracy indicators such as MCC change depending on the number of applications, and an example is shown where the MCC is highest under the condition of 25 applications. In other words, data augmentation does not always improve accuracy when applied, and there may be an optimal range for the number of applications (number of transfers). Therefore, a configuration in which the data augmentation unit 150 determines the number of transfers (transfer amount) for the purpose of maximizing accuracy (user specification or automatic search), as in the second embodiment, is effective in efficiently obtaining accuracy improvements.

[0134] Figure 15 shows an example where Precision / Recall / F1 under the "interruption present" condition changes depending on whether data augmentation is used and the number of times it is applied. In other words, by appropriately applying data augmentation under conditions that include a small class of interruptions, interruption detection / classification performance (e.g., F1 and MCC) can be improved. This can be expected to reduce the likelihood of overlooking interruptions (improve recall) and improve overall classification performance.

[0135] As shown in Figure 15, while LSTM + Shap can achieve a certain level of accuracy, it is shown to have a long factor calculation time. In contrast, the configuration of the second embodiment (for example, a combination of attention-based estimation and data augmentation) is shown to be able to achieve equivalent accuracy while keeping the calculation time down. In other words, when factor estimation is repeatedly performed in operation, it can be advantageous in terms of computational resources and processing time.

[0136] Based on the above, according to the second embodiment, the data expansion unit 150 generates training data by transferring interruption intervals extracted from the field log to other intervals, and determines the number of transfers (transfer amount) from the viewpoint of maximizing accuracy, thereby (1) improving the accuracy of interruption detection by mitigating insufficient learning caused by small amounts of unbalanced data, (2) stabilizing the factor estimation results (reducing noise), and (3) achieving the desired accuracy while suppressing calculation time depending on the method. In other words, supervised learning can be easily applied to actual operations even for video interruptions that are rare in the field.

[0137] (Example of using the data expansion unit 150 as a standalone device) In the second embodiment, the data expansion unit 150 may be configured as a functional block of the information processing device 100, but is not limited thereto. That is, the data expansion unit 150 may be used as a standalone device (referred to as the information processing device 200).

[0138] Figure 16 is a diagram showing the configuration of the information processing device 200. As shown in Figure 16, the information processing device 200 includes an input unit 151, a data expansion unit 150, and an output unit 152. Note that the configuration shown in Figure 16 can be considered as a detailed configuration of the "data expansion unit 150" shown in Figure 12.

[0139] The input unit 151 acquires training data to be augmented. Specifically, the input unit 151 receives time-series data (multiple explanatory variables) such as logs and measurements acquired from the system to be factor-estimated, and a target variable (for example, a label indicating whether or not there is a video interruption) associated with the time-series data. In other words, the input unit 151 takes in the dataset (explanatory variable column and target variable column) that will be input to the data augmentation process in a predetermined format (for example, tabular format, time-series array format, etc.) and passes it to the data augmentation unit 150.

[0140] The data augmentation unit 150 performs data augmentation processing on the dataset acquired by the input unit 151 to supplement small data classes (e.g., interruptions), and generates an augmented dataset. The specific method of data augmentation in the data augmentation unit 150 is as described above.

[0141] The output unit 152 outputs the expanded dataset generated by the data expansion unit 150 to the outside. Specifically, the output unit 152 outputs the expanded explanatory variable sequence and the expanded target variable sequence in a format that can be used as training data. In other words, the output unit 152 supplies the expanded data to the preprocessing unit 110 (or inference unit 120) of the information processing device 100, enabling supervised learning and evaluation based on the expanded data.

[0142] Furthermore, the output unit 152 may transmit the expanded data via a communication interface, or it may save it to a storage medium. In other words, the output unit 152 can select an appropriate output format depending on the computing environment (cloud, edge server, terminal, etc.) in which the subsequent learning process is performed.

[0143] (Example Hardware Configuration) The devices (information processing device 100, information processing device 200) described in this embodiment (first and second embodiment) can be realized, for example, by having a computer execute a program. This computer may be a physical computer or a virtual machine on the cloud.

[0144] In other words, the device can be realized by using hardware resources such as the CPU and memory built into a computer to execute a program corresponding to the processing performed by the device. The program can be recorded on a computer-readable recording medium (such as portable memory), saved, and distributed. It can also be provided via a network, such as the Internet or email.

[0145] Figure 17 shows an example of the hardware configuration of the computer described above. The computer in Figure 17 has a drive device 1000, an auxiliary storage device 1002, a memory device 1003, a CPU 1004, an interface device 1005, a display device 1006, an input device 1007, an output device 1008, etc., all of which are interconnected by bus B. The computer may also be equipped with a GPU.

[0146] The program that enables processing on the computer is provided on a recording medium 1001, such as a CD-ROM or memory card. When the recording medium 1001 containing the program is set in the drive device 1000, the program is installed from the recording medium 1001 to the auxiliary storage device 1002 via the drive device 1000. However, the program does not necessarily have to be installed from the recording medium 1001; it may also be downloaded from another computer via a network. The auxiliary storage device 1002 stores the installed program as well as necessary files and data.

[0147] The memory device 1003 reads and stores a program from the auxiliary storage device 1002 when a program startup command is received. The CPU 1004 implements the functions related to the memory device 1003 according to the program stored in the memory device 1003. The interface device 1005 is used as an interface for connecting to a network, etc. The display device 1006 displays a GUI (Graphical User Interface) etc., based on a program. The input device 1007 consists of a keyboard and mouse, buttons, or a touch panel, etc., and is used to input various operation commands. The output device 1008 outputs the calculation results.

[0148] Furthermore, the functions of the elements disclosed herein may be implemented using circuits or processing circuitry that include general-purpose processors, special-purpose processors, integrated circuits, ASICs (Application Specific Integrated Circuits), FPGAs (Field Programmable Gate Arrays), conventional circuits, and / or combinations thereof that are programmed using one or more programs stored in one or more memories, or otherwise configured to perform the disclosed functions. A processor is considered processing circuitry or circuitry because it includes transistors and other circuits. A processor may be a programmed processor that executes programs stored in memory. In this disclosure, a circuit, unit, or means is hardware that performs the enumerated functions, or hardware programmed to perform the enumerated functions. Hardware may be any hardware disclosed herein that is programmed or configured to perform the enumerated functions.

[0149] The system includes memory for storing computer programs, which include computer instructions. These computer instructions provide logic and routines that enable hardware (e.g., processing circuitry or circuitry) to perform the methods disclosed herein. The computer programs can be implemented in commonly known forms, such as computer-readable storage media, computer program products, memory devices, recording media such as CD-ROMs and DVDs, and / or memory in FPGAs and ASICs.

[0150] (Summary of Embodiments, Effects, etc.) As described above, the technology described in these embodiments (first and second embodiments) makes it possible to generate new vectors that capture the changes in the relationships between features over time, and by constructing a machine learning model using these vectors, it becomes possible to analyze the correct factors based on the relationships between features.

[0151] Furthermore, by using a time-series model as the machine learning model, it becomes possible to analyze logs while considering their time-series nature, thus addressing cases where conventional techniques could not effectively analyze the contributing factors.

[0152] The following additional information is disclosed regarding the embodiments described above.

[0153] <Note> (Note 1) An information processing device comprising: a first processing unit that acquires multiple feature quantities that change over time from a communication system and generates relational feature quantities that express the relationship between feature quantities for each combination of two feature quantities; a second processing unit that estimates whether or not a predetermined event has occurred based on the relational feature quantities; and a third processing unit that estimates the factors of the predetermined event based on one or more relational feature quantities that formed the basis of the estimation result of whether or not the predetermined event has occurred.

[0154] (Appendix 2) The information processing apparatus according to Appendix 1, wherein the first processing unit acquires a temporal subvector as the feature quantity and determines whether or not there is a relationship between the subvectors, thereby generating the relational feature quantity as a new vector.

[0155] (Note 3) The second processing unit is an information processing device according to Note 1, which estimates whether or not the predetermined event has occurred using a machine learning model suitable for processing time-series data.

[0156] (Appendix 4) The information processing apparatus according to Appendix 1, wherein the third processing unit calculates the importance of each related feature for the estimation result of whether the predetermined event has occurred, and estimates the factors of the predetermined event based on one or more related features whose importance satisfies predetermined conditions.

[0157] (Appendix 5) The predetermined condition is that the importance is greater than or equal to a threshold which is the average of the importance in negative examples, and the third processing unit estimates the factors of the predetermined event by combining the same source feature quantities in at least two relational feature quantities that satisfy the predetermined condition, as described in Appendix 4.

[0158] (Note 6) An information processing device comprising: an input unit that inputs learning data including time-series data, which includes a first class interval corresponding to the occurrence of a predetermined event and a second class interval that is an interval other than the first class interval; and a data extension unit that obtains first data which is data included in the first class interval from the learning data, and transfers the first data to the second class interval to generate extended learning data which is an extension of the learning data.

[0159] (Addendum 7) The information processing apparatus according to Addendum 6, wherein the first data consists of a plurality of data that are consecutive in time series, and the data expansion unit transfers the consecutive data to the second class interval.

[0160] (Appendix 8) The data augmentation unit determines the number of transfers or the amount of transfers by user specification or by searching using an evaluation index based on learning results using the augmented learning data, as described in Appendix 6.

[0161] (Appendix 9) A factor estimation method executed by an information processing device, comprising: a step of acquiring a plurality of feature quantities that change over time from a communication system, and for each combination of two feature quantities, generating a relational feature quantity that expresses the relationship between the feature quantities; a step of estimating whether or not a predetermined event has occurred based on the relational feature quantity; and a step of estimating the factors of the predetermined event based on one or more relational feature quantities that formed the basis of the estimation result of whether or not the predetermined event has occurred.

[0162] (Note 10) A method for generating learning data to be executed by an information processing device, comprising: the step of inputting learning data which includes time-series data and includes a first class interval corresponding to the occurrence of a predetermined event and a second class interval which is an interval other than the first class interval; and the step of obtaining first data which is data included in the first class interval from the learning data, and transferring the first data to the second class interval to generate extended learning data which is an extension of the learning data.

[0163] (Appendix 11) A non-temporary storage medium storing a program for causing a computer to function as a part of an information processing device described in any one of the appendix items 1 through 8.

[0164] Although this embodiment has been described above, the present invention is not limited to this specific embodiment, and various modifications and changes are possible within the scope of the gist of the invention as described in the claims.

[0165] This patent application claims priority based on the international patent application PCT / JP2025 / 006108 filed on February 21, 2025, and the entire contents of the international patent application PCT / JP2025 / 006108 are incorporated herein by reference.

[0166] 100, 200 Information Processing Device 110 Pre-processing Unit 120 Estimation Unit 130 Importance Calculation Unit 140 Post-processing Unit 150 Data Expansion Unit 151 Input Unit 152 Output Unit 1000 Drive Device 1001 Recording Medium 1002 Auxiliary Storage Device 1003 Memory Device 1004 CPU 1005 Interface Device 1006 Display Device 1007 Input Device 1008 Output Device

Claims

1. An information processing device comprising: a first processing unit that acquires multiple feature quantities that change over time from a communication system and generates relational feature quantities that express the relationship between the feature quantities for each combination of two feature quantities; a second processing unit that estimates whether or not a predetermined event has occurred based on the relational feature quantities; and a third processing unit that estimates the factors of the predetermined event based on one or more relational feature quantities that formed the basis of the estimation result of whether or not the predetermined event has occurred.

2. The information processing apparatus according to claim 1, wherein the first processing unit acquires a temporal subvector as the feature quantity, determines whether or not there is a relationship between the subvectors, and generates the relational feature quantity as a new vector.

3. The information processing apparatus according to claim 1, wherein the second processing unit estimates whether or not the predetermined event has occurred using a machine learning model suitable for processing time-series data.

4. The information processing apparatus according to claim 1, wherein the third processing unit calculates the importance of each related feature for the estimation result of whether the predetermined event has occurred, and estimates the factors of the predetermined event based on one or more related features whose importance satisfies predetermined conditions.

5. The information processing apparatus according to claim 4, wherein the predetermined condition is that the importance is greater than or equal to a threshold which is the average value of the importance in negative examples, and the third processing unit estimates the factors of the predetermined event by combining the same source feature quantities in at least two relational feature quantities that satisfy the predetermined condition.

6. An information processing device comprising: an input unit that inputs learning data including time-series data, which includes a first class interval corresponding to the occurrence of a predetermined event and a second class interval that is an interval other than the first class interval; and a data extension unit that obtains first data from the learning data, which is data included in the first class interval, and transfers the first data to the second class interval, thereby generating extended learning data which is an extension of the learning data.

7. The information processing apparatus according to claim 6, wherein, if the first data consists of a plurality of data that are consecutive in time series, the data expansion unit transfers the consecutive data together to the second class interval.

8. The information processing apparatus according to claim 6, wherein the data augmentation unit determines the number of transfers or the amount of transfers by user specification or by searching using an evaluation index based on learning results using the augmented learning data.

9. A factor estimation method executed by an information processing device, comprising: a step of acquiring a number of time-series changing feature quantities from a communication system, and for each combination of two feature quantities, generating a relational feature quantity that expresses the relationship between the feature quantities; a step of estimating whether or not a predetermined event has occurred based on the relational feature quantity; and a step of estimating the factors of the predetermined event based on one or more relational feature quantities that formed the basis of the estimation result of whether or not the predetermined event has occurred.

10. A method for generating learning data executed by an information processing device, comprising: inputting learning data which includes time-series data and includes a first class interval corresponding to the occurrence of a predetermined event and a second class interval which is an interval other than the first class interval; and generating extended learning data which is an extension of the learning data by obtaining first data which is data included in the first class interval from the learning data and transferring the first data to the second class interval.

11. A program for causing a computer to function as a component of an information processing device described in any one of claims 1 to 8.

Images