Information processing device, information processing method, and computer program

Abstract

This makes it easier to identify variables that influence the error or goodness-of-fit indicator between the predicted value from the predictive model and the correct value. [Solution] The information processing device includes a processing unit. The processing unit receives multiple first time series data for each of multiple input variables in a first period and constructs a prediction model that predicts one or more output variables at times after the first period. The processing unit calculates third time series data of an index that represents the error or goodness of fit between one or more second time series data representing one or more output variables predicted by the prediction model for each of the one or more time points included in the first period, and ground truth time series data representing the correct answers for one or more output variables at one or more time points included in the first period. The processing unit uses the multiple first time series data and the third time series data to construct a time series causal graph that represents the causal relationship between multiple input variables and the index.

Description

[Technical Field]

[0001] Embodiments of the present invention relate to an information processing apparatus, an information processing method, and a computer program. [Background technology]

[0002] Predictive methods using machine learning-based predictive models demonstrate remarkable performance across a wide range of prediction tasks. However, the black-box nature of these predictive models can pose a problem when applying such methods to socially impactful infrastructure services.

[0003] As an attempt to provide interpretations for predictive models that have a black-box nature, methods such as SHAP (SHapley Additive exPlanations) and LIME (Local Interpretable Model-agnostic Explanations) are known, which attempt to provide ex-post explanations of predictive models. These methods are intended to explain what causes the predicted values ​​given the input data. In practice, however, there is often more interest in why the predicted values ​​deviate from the actual or measured values ​​that represent the correct answer, rather than in interpreting the predictive model itself. [Prior art documents] [Patent Documents]

[0004] [Patent Document 1] Japanese Patent Publication No. 2023-067749 [Non-patent literature]

[0005] [Non-Patent Document 1] Ide, T., Dhurandhar, A., Navratil, J., Singh, M., & Abe, N., “Anomaly Attribution with Likelihood Compensation.”, Proceedings of the AAAI Conference on Artificial Intelligence, 35(5), 4131-4138,2021. [Non-Patent Document 2] Kailash Budhathoki, Lenon Minorics, Patrick Bloebaum, Dominik Janzing, “Causal structure-based root cause analysis of outliers.”, Proceedings of the 39th International Conference on Machine Learning, PMLR 162:2357-2369, 2022. [Non-Patent Document 3] J. Runge, P. Nowack, M. Kretschmer, S. Flaxman, D. Sejdinovic, “Detecting and quantifying causal associations in large nonlinear time series datasets.”, Sci. Adv. 5, eaau4996, 2019. [Overview of the project] [Problems that the invention aims to solve]

[0006] The present invention aims to provide an information processing device, an information processing method, and a computer program that can more easily identify variables that affect the error between predicted values ​​from a prediction model and values ​​representing the correct answer, or an index representing the degree of fit. [Means for solving the problem]

[0007] The information processing device of the embodiment includes a processing unit. The processing unit receives multiple first time series data for each of multiple input variables in a first period and constructs a prediction model that predicts one or more output variables at times after the first period. The processing unit calculates third time series data of an index that represents the error or goodness of fit between one or more second time series data representing one or more output variables predicted by the prediction model for each of the one or more time points included in the first period, and ground truth time series data representing the correct answers for one or more output variables at one or more time points included in the first period. The processing unit uses the multiple first time series data and the third time series data to construct a time series causal graph that represents the causal relationship between multiple input variables and the index. [Brief explanation of the drawing]

[0008] [Figure 1] Block diagram of an information processing device according to an embodiment. [Figure 2] Flowchart of the analysis process in the embodiment. [Figure 3] A diagram showing an example of combined time series data. [Figure 4] A diagram illustrating an example of a time-series causal graph. [Figure 5] A diagram showing an example of training data. [Figure 6] Hardware configuration diagram of the information processing device according to the embodiment. [Modes for carrying out the invention]

[0009] A preferred embodiment of the information processing device according to this invention will be described in detail below with reference to the attached drawings.

[0010] In the following embodiments, a predictive model EM (predictive model) is used that predicts one or more output variables (dependent variables) at a specified time (e.g., a future time) based on multiple time series data TDA (first time series data), which are the time series data of multiple input variables (explanatory variables). Below, we will mainly describe an example of a predictive model EM that predicts one output variable, but the same procedure can be applied to a predictive model EM that predicts multiple output variables. Note that prediction is not limited to estimating information (such as output variables) at a future time, but includes estimating information at any time.

[0011] Multiple input variables may include variables of the same type as the variable to be predicted (output variable) and obtained before a specified time. In other words, the prediction model EM may be a model that predicts the value of a variable at a future time as the output variable, based on multiple input variables that include the variable's past values.

[0012] Time series data (TDA) is data collected from various devices used in the target system, for example. The target system can be any system, but examples include the following systems: This system takes observed rainfall data from multiple locations as input variables and predicts the amount of water inflow from rainwater at a specified location (dam, river, etc.) as the output variable. This system uses observed values ​​such as wind speed and relative humidity from multiple locations as input variables to predict visibility at a specified location (airport, road, etc.), which is affected by fog, blizzards, etc., as the output variable. A system that uses time-series data (sensor values, control values, etc.) collected from monitored systems such as plants as input variables to predict (detect) output variables indicating abnormalities in the monitored systems.

[0013] As described above, in predictions using the machine learning-based prediction model EM, there is often interest in why the predicted values ​​deviate from the actual values. For example, it may be necessary to identify the variables that influence the deviation between the actual values ​​and the predicted values. This deviation can be represented by, for example, the error (prediction error), which is the difference between the actual values ​​and the predicted values. The indicator used to represent the deviation is not limited to the error; any other indicator may be used. For example, the degree of fit, which indicates the degree to which the actual values ​​and predicted values ​​match, may be used. Below, we will mainly explain examples in which the error is used as an indicator to represent the deviation.

[0014] For example, consider a system that predicts dam inflow. Suppose the prediction model EM, while generally having a good prediction error, exhibits a situation where, despite significant rainfall and high actual inflow, it predicts a lower inflow (a discrepancy between actual and predicted values). If the reason for this discrepancy cannot be explained, it becomes difficult to adopt the EM prediction model.

[0015] Techniques have been proposed to identify the extent to which each input variable contributes to the judgment that the discrepancy between actual and predicted values ​​in given test data is abnormal. However, such techniques calculate the "degree of responsibility" of statistically related variables and do not consider the causal relationships between variables.

[0016] Therefore, the following embodiment identifies variables that influence the discrepancy between actual and predicted values ​​for given test data, taking into account the causal relationships between variables in prediction using the prediction model EM.

[0017] The embodiment includes at least some of the following functions, for example: • Predicts the value of an output variable at a specified future time based on multiple historical time-series data. • The difference between the predicted and actual values ​​of the output variable at past points in time is saved as the prediction error. • Explore time-series causal graphs that show the causal relationship between past prediction errors and multiple historical time-series data. • Construct a generative model (GM) that corresponds to time-series causal graphs. • Using the generative model GM, calculate the contribution of each time series data (variable) to the prediction error at a specified time.

[0018] In this embodiment, a structural causal model (time-series causal graph) is constructed between multiple variables, with the prediction error added as a single variable in addition to the multiple variables used for prediction. This makes it possible to plot the causal relationship between the prediction error and its root cause as a graph, making the interpretation of the prediction model easier.

[0019] Furthermore, in this embodiment, a generative model GM is constructed, which is a model that generates predicted values ​​of time series data corresponding to each node in a time series causal graph. Using such a generative model GM, it is possible to calculate the contribution of variables that are candidates for the root cause of the prediction error.

[0020] In some cases, the input variables include not only the actual values ​​(such as observed values) of a variable IDA (the first variable), but also predicted values ​​(such as weather forecast values) of variable IDA. In such cases, the difference between the actual and predicted values ​​of variable IDA may also be set as a new variable, and a time-series causal graph between variables can be constructed. This makes it possible to analyze the relationship between the prediction error by the prediction model EM and the difference with respect to variable IDA.

[0021] Figure 1 is a block diagram showing an example of the configuration of an information processing device 100 according to an embodiment. As shown in Figure 1, the information processing device 100 includes a storage unit 131, an acquisition unit 101, a prediction model construction unit 111, an indicator calculation unit 112, a coupling unit 113, a graph construction unit 114, a generation model construction unit 121, a contribution calculation unit 122, and an output control unit 102.

[0022] The storage unit 131 stores various types of information used by the information processing device 100. For example, the storage unit 131 stores various types of information acquired by the acquisition unit 101 (such as time-series data), as well as processing results from each unit.

[0023] The storage unit 131 can be composed of any commonly used storage medium, such as flash memory, memory card, RAM (Random Access Memory), HDD (Hard Disk Drive), and optical disc.

[0024] The acquisition unit 101 acquires various types of information used by the information processing device 100. For example, the acquisition unit 101 acquires multiple time series data TDAs used for prediction by the prediction model EM. These multiple time series data TDAs are, for example, data that have been collected in advance and stored in an external database or similar device outside the information processing device 100.

[0025] A physical relationship can exist between multiple time series data (TDAs). For example, in the case of multiple time series data (TDAs) obtained by measuring the flow rate of a dam or river or rainfall, the sum of the flow rates or rainfall from the upstream dam or river can be considered to correspond to the flow rate of the downstream dam or river. This is because the water from the upstream dam or river, and the rain that falls in the river basin, flow into the downstream dam or river. Therefore, for example, there is a physical relationship between the time series data of an input variable representing the flow rate of an upstream dam and the time series data of an input variable representing the flow rate of a downstream dam or river.

[0026] As described above, multiple time series data TDA may include not only time series data of actual values ​​of a variable IDA, but also time series data of predicted values ​​(such as weather forecast values) of the variable IDA. For example, multiple time series data TDA may include time series data representing actual rainfall at a certain location and time series data representing predicted rainfall at the same location.

[0027] The method by which the acquisition unit 101 acquires information can be any method, but for example, it can be a method of receiving information from an external device via a network, or a method of reading information from a storage medium.

[0028] The prediction model construction unit 111 constructs a prediction model EM. For example, the prediction model construction unit 111 takes multiple time series data TDA from a certain period TA (first period) as input and constructs a prediction model EM that predicts the output variable at a time after period TA.

[0029] The indicator calculation unit 112 calculates time series data of the prediction error (an example of an indicator), which is the difference between the predicted value by the prediction model EM and the actual value. For example, the indicator calculation unit 112 calculates time series data TDC (third time series data) of the error between one or more time series data TDB (second time series data) and the ground truth time series data.

[0030] One or more time series data TDBs correspond to one or more time series data representing the output variables predicted by the prediction model EM for each of the one or more time points included in period TA. The ground truth time series data represents the correct values ​​of the output variables at one or more time points included in period TA. For example, the ground truth time series data corresponds to the actual values ​​of the variables corresponding to the output variables obtained at each time point included in period TA.

[0031] The merging unit 113 generates combined time series data by combining multiple time series data TDA and time series data TDC.

[0032] The graph construction unit 114 constructs a time-series causal graph using multiple time-series data TDA and time-series data TDC. For example, the graph construction unit 114 constructs a time-series causal graph using combined time-series data. The time-series causal graph corresponds to a graph that represents the causal relationship between multiple input variables (multiple time-series data TDA) and the prediction error (time-series data TDC).

[0033] If time series data of the predicted value of variable IDA is available, the graph construction unit 114 may further use the time series data TDD (fourth time series data) of the difference between the actual value of variable IDA and the predicted value of variable IDA to construct a time series causal graph. For example, the merging unit 113 generates combined time series data by merging multiple time series data TDA, time series data TDC, and time series data TDD. The graph construction unit 114 constructs a time series causal graph using the combined time series data including the time series data TDD.

[0034] The generative model construction unit 121 constructs a generative model GM using a time-series causal graph. The generative model corresponds to a model that generates input variables corresponding to node NB (second node), which is a child node of node NA (first node), from input variables corresponding to node NA (first node) included in the time-series causal graph.

[0035] The contribution calculation unit 122 uses a generative model to calculate the contribution of multiple input variables to the prediction error.

[0036] The output control unit 102 controls the output of various types of information used by the information processing device 100. For example, the output control unit 102 outputs the constructed time-series causal graph and the calculated contribution. The method of outputting the information can be any method, but for example, it can be displayed on a display device or transmitted to an external device via a network.

[0037] At least a portion of each of the above parts (acquisition unit 101, prediction model construction unit 111, indicator calculation unit 112, coupling unit 113, graph construction unit 114, generative model construction unit 121, contribution calculation unit 122, and output control unit 102) may be implemented by one or more processing units. Each of the above parts may be implemented by, for example, one or more processors. For example, each of the above parts may be implemented by having a processor such as a CPU (Central Processing Unit) and a GPU (Graphics Processing Unit) execute a program, i.e., by software. Each of the above parts may be implemented by a processor such as a dedicated IC (Integrated Circuit), i.e., by hardware. Each of the above parts may be implemented by using both software and hardware. When multiple processors are used, each processor may implement one of the above parts, or two or more of the above parts.

[0038] Furthermore, the information processing device 100 may be composed of one physical device or multiple physical devices. For example, the information processing device 100 may be built on a cloud environment. Also, each part of the information processing device 100 may be distributed and provided on multiple devices.

[0039] Next, the analysis process performed by the information processing device 100 of this embodiment will be described. Figure 2 is a flowchart showing an example of the analysis process in this embodiment. The analysis process includes processes for constructing a time-series causal graph and calculating the contribution of variables to the prediction error.

[0040] First, the prediction model construction unit 111 constructs a prediction model EM using multiple time series data acquired by the acquisition unit 101 (step S101). The prediction model EM is a model that takes time series data of a specified input variable from the multiple acquired time series data as input, predicts the value of a specified output variable, and outputs it. For example, the prediction model construction unit 111 constructs the prediction model EM using multiple time series data TDA obtained at multiple past time points as training data. In the following, the period over which training data is obtained may be referred to as the training period.

[0041] For example, when predicting the inflow to a specific dam, multiple time series data, such as time series data of the dam's inflow for the past several years and time series data of rainfall for the past several years (or time series data of past predicted rainfall), are used as training data. The prediction model construction unit 111 learns (constructs) a prediction model EM for predicting the dam's inflow at a specified time in the future, based on the relationship between the dam's inflow included in the training data and multiple time series data such as rainfall or predicted rainfall.

[0042] Let's further explain the training of the prediction model EM. The training period is defined as time 1, ..., T1. The test period is defined as time T1+1, ..., T1+T2. The test period corresponds to the period during which time series data used in the process of calculating the contribution of each variable to the prediction error is obtained. For the sake of explanation, we assume that the test period starts immediately after the training period (T1+1). The test period can start at any time after the training period. That is, the same procedure can be applied even if the test period starts at time T1+h (where h is a positive value).

[0043] The following equation (1) represents the time series data of multiple (e.g., p) time series during the training period. The following equation (2) represents the time series data of the variable corresponding to the target of prediction (output variable) during the training period.

number

number

[0044] As a result, the training data D1 is represented by equation (3) below. The training data D1 corresponds to multiple time series data TDA.

number

[0045] Multiple (e.g., p) time series data during the test period are represented by equation (4) below. The time series data of the variable corresponding to the target of prediction (output variable) during the test period are represented by equation (5) below.

number

number

[0046] Therefore, the test data D2 can be expressed by the following equation (6).

number

[0047] Each time point included in the training and test periods is within a fixed time interval. This fixed time interval can be any value, for example, a 1-day interval or a 1-hour interval. The length of the training period T1 and the length of the test period T2 can be any value, for example, T1 is 5 years and T2 is 1 year.

[0048] The prediction model building unit 111 uses the training data D1 to obtain time series data X of length w of the input variable up to time t ∈ {1, ..., T1}. 1,(t-w+1):t ,X 2,(t-w+1):t ,···,X p,(t-w+1):t ,Y (t-w+1):t And the output variable Y at time (t+1) t+1The relationship with is modeled as in the following equation (7).

Number

[0049] X i,(t-w+1):t is (X i,t-w+1 , X i,t-w+2 , ···, X i,t-1 , X i,t ). Y (t-w+1):t is (Y t-w+1 , Y t-w+2 , ···, Y t-1 , Y t ). ε represents the error term.

[0050] In the example of equation (7), the function f(·) corresponds to the prediction model EM. The function f(·) can be any model as long as it inputs time series data of multiple input variables and outputs predicted values of one or more output variables. For example, the function f(·) can be a simple model such as a linear regression model or a neural network model such as LSTM (Long Short Term Memory).

[0051] The learning method of the function f(·) using the training data D1 can be any method applicable to the function f(·) adopted. For example, the prediction model construction unit 111 can use the (T1 - w) length - w (p + 1) - variable time series data represented by the following equation (8) included in the training data D1, and obtain the parameters of the function f that minimize a loss function such as the squared loss represented by the following equation (9) by an optimization method such as the stochastic gradient method.

Number

Number

[0052] The predicted value of the function f based on the parameters learned in this way is represented by f^ (f - hat: a symbol with a hat symbol attached above the symbol f).

[0053] The loss function to be minimized is not limited to the squared loss, but can be any other function representing any loss. For example, the loss function may be an L1 loss as shown in equation (10) below.

number

[0054] Furthermore, instead of finding the parameters that minimize the loss function, the prediction model construction unit 111 may find the parameters of the function f that maximize the index equivalent to goodness of fit using an optimization method such as the stochastic gradient method. The index equivalent to goodness of fit is, for example, the error ε expressed by equation (11) below. t This is the log-likelihood assuming that follows a specific probability distribution. A specific probability distribution is, for example, the mean μ expressed by equation (12) and the variance σ expressed by equation (13). 2 This is a normal distribution. The log-likelihood in this case is given by equation (14) below. Goodness-of-fit measures such as the log-likelihood can be used as indicators of deviation instead of the prediction error as described above.

number

number

number

number

[0055] Returning to the explanation of Figure 2, once the prediction model EM (function f^) is constructed as described above, the index calculation unit 112 calculates the error (prediction error) between the predicted value and the actual value using the prediction model EM (step S102). For example, the index calculation unit 112 uses the function f^ to calculate the prediction error R at time t ∈ {w,···,T1-1}. t+1The following equation (15) is used to calculate the predicted value of the output variable. The index calculation unit 112 obtains the predicted value of the output variable by inputting the time series data of length w contained in the training data D1 into the function f^. In (14), Y t+1 This corresponds to the actual value of the output variable at time (t+1).

number

[0056] The prediction error is not limited to a simple difference as shown in equation (15), but may also be the absolute value of the difference, or the square of the difference.

[0057] Returning to the explanation of Figure 2, the merging unit 113 generates a combined time series data by combining multiple time series data TDA and time series data TDC, which corresponds to the time series data of the prediction error (step S103). Equation (16) below shows an example of each time series data that makes up the combined time series data. For example, the prediction error R (w+1):T_1 This is the prediction error R calculated in step S102 for each time t ∈ {w,···,T1-1}. t+1 This represents the time series data TDC of the error. i,(w+1):T_1 , and, Y (w+1):T_1 This corresponds to the time series data TDA for the same period (time w to time T1-1).

number

[0058] The coupling portion 113 has a prediction error R represented by equation (16). (w+1):T_1 And, X i,(w+1):T_1 And, Y (w+1):T_1 The two are combined to generate the combined time series data represented by equation (17).

number

[0059] Combined time series data can be viewed as tabular data, as shown in Figure 3. Figure 3 is an example of combined time series data represented in tabular format.

[0060] In Figure 3, the first column represents time, the second to p+1 columns represent p time series data, the p+2 column represents time series data of the variable corresponding to the prediction target (output variable), and the p+3 column represents time series data of the prediction error calculated by the index calculation unit 112.

[0061] Returning to the explanation of Figure 2, the graph construction unit 114 constructs a time-series causal graph for the generated combined time-series data (step S104). For example, the graph construction unit 114 uses the generated combined time-series data to construct a time-series causal graph that represents the causal relationships of the time-series data between the second column and the p+3th column. The graph construction unit 114 can construct a time-series causal graph from time-series data by, for example, using the PCMCI algorithm (e.g., Non-Patent Document 3).

[0062] Figure 4 shows an example of a constructed time-series causal graph. Note that Figure 4 is an example of a time-series causal graph when w=2. The circles in Figure 4 represent nodes that make up the graph. In the example in Figure 4, the time-series causal graph includes nodes corresponding to the following variables. X 1,t-2 ,X 2,t-2 ,···,X p,t-2 ,Y t-2 ,R t-2 , X 1,t-1 ,X 2,t-1 ,···,X p,t-1 ,Y t-1 ,R t-1 , X 1,t ,X 2,t ,···,X p,t ,Y t ,R t

[0063] Returning to the explanation of Figure 2, the generative model construction unit 121 constructs (learns) a generative model corresponding to the time-series causal graph constructed by the graph construction unit 114 (step S105). The generative model is a model that corresponds to the data generation mechanism of the time-series components corresponding to each node of the time-series causal graph.

[0064] For example, the generative model construction unit 121 assumes that the time series components corresponding to each node are generated by a function like equation (18) below.

number

[0065] pa(·) represents the parent node in the time-series causal graph. i,t is the variable X i,t This represents noise in the generation of N. y,t is the variable Y t This represents noise in the generation of g. X_i,t (·) is X i,t This represents a function that generates g. Y_t (·) is Y t This represents a function that generates g. R_t (·)=Y t -f^(·) is R t This represents a function that generates [something].

[0066] For example, the first line of equation (18) is the input variable X i,t This corresponds to a generative model for the input variable Y. t This corresponds to a generative model for (18). The generative model in equation (18) corresponds to a model that calculates a value for the variable corresponding to the child node by adding noise to a value calculated based on the value of the variable corresponding to the parent node. The third line can be expressed using the already obtained prediction model EM(function f^). Therefore, g R_t (·) does not need to be predicted.

[0067] The generative model construction unit 121, for example, the function g X_i,t (·), and g Y_t Assuming (·) is linear, gX_i,t (·), and g Y_t Predict (·). The predicted function is indicated by the hat symbol "^" and g X_i,t ^(·) and g Y_t It is sometimes represented as ^(·).

[0068] First, the generative model construction unit 121 generates the combined time series data shown in equation (19) below, with the time range extended to T1+T2.

number

[0069] The generative model construction unit 121 uses the generated combined time series data to construct the time series data X corresponding to each node in the following time series causal graph. 1,t ~X p,t ,Y t ,X 1,t-1 ~X p,t-1 ,···,X 1,t-w ~X p,t-w ,···,R t This generates the following. This time-series data can be interpreted as data delayed up to T2, and therefore may be referred to as time-delayed data below. X 1,t ,X 2,t ,···,X p,t ,Y t ,R t , X 1,t-1 ,X 2,t-1 ,···,X p,t-1 ,Y t-1 ,R t-1 ,···, X 1,t-w ,X 2,t-w ,···,X p,t-w ,Y t-w ,R t-w

[0070] The generative model construction unit 121, for example, uses at least a portion of the time-delayed data as training data to construct a linear function g X_i,t (·) and function g Y_tLearn (·). FIG. 5 is a diagram showing an example of training data used in this case. As shown in FIG. 5, the training data can be regarded as tabular data. For example, the generation model construction unit 121 uses the time-series data of the column corresponding to pa(X i,t ) in the training data of FIG. 5 to learn the function g X_i,t (·). Also, the generation model construction unit 121 uses the time-series data of the column corresponding to pa(Y t ) in the training data of FIG. 5 to learn the function g Y_t (·).

[0071] Return to the description of FIG. 2. The contribution degree calculation unit 122 calculates the contribution degree of each variable to the prediction error R t+1 at a specified time using the generation model (step S106). The prediction error R t+1 is the prediction error at a specified time t + 1 ∈ {T1 + 1, ···, T1 + T2} and is represented by the following equation (20).

Equation

[0072] In the case of the example of the time-series causal graph in FIG. 4, the contribution degree of each variable X 1,t , X 2,t , ···, X p,t , Y t+1 to the prediction error R t+1 is obtained. For example, the contribution degree calculation unit 122 calculates the contribution degree by a method (for example, Non-Patent Document 2) of calculating the contribution degree based on the Shapley value from the generation model.

[0073] Specifically, as the contribution degree Φ(j) of each node j included in the node set V shown in the following equation (21) to R t , the Shapley value as shown in the following equation (22) can be used.

Equation

Equation

[0074] C(j|I) is expressed by equation (23) below. S(Rt)^rd(I∪{j}) is expressed by equation (24). S(Rt)^rd(I) is expressed by equation (25).

number

number

number

[0075] Equation (24) corresponds to the value obtained by applying the following procedure. For i∈I∪{j}, g X_i,t ^(·) and g Y_t Using ^(·), (T1+T2-w) noises are calculated as shown in equation (26) below. By sampling with overlap, (T1+T2-w) noises are generated again.

number

[0076] Equation (25) corresponds to the value obtained by applying the same procedure as above to i∈I instead of i∈I∪{j}.

[0077] Furthermore, in situations where the number of nodes exceeds several tens, the computational cost of calculating the contribution Φ(j) can become substantial. In such cases, a Monte Carlo approximation may be used, where the set I appearing in the formula for contribution Φ(j) is sampled using the Monte Carlo method, and the resulting value of contribution Φ(j) is used.

[0078] Through the analysis process described above, it is possible to identify which points in time the prediction error—the discrepancy between the actual and predicted values ​​of the output variable being predicted—causally depends on the values ​​of the input and output variables. As a result, it becomes possible to interpret the results of the prediction model and to provide a guideline for which input variables need to be improved to reduce the prediction error.

[0079] Thus, in the information processing device of this embodiment, it is possible to more easily identify variables that affect the error between the predicted value from the prediction model and the value representing the correct answer, or the index representing the degree of fit.

[0080] Next, the hardware configuration of the information processing device of the embodiment will be described using Figure 6. Figure 6 is an explanatory diagram showing an example of the hardware configuration of the information processing device of the embodiment.

[0081] The information processing device of this embodiment includes a control device such as a CPU (Central Processing Unit) 51, a storage device such as a ROM (Read Only Memory) 52 and a RAM (Random Access Memory) 53, a communication interface 54 that connects to a network for communication, and a bus 61 that connects each part.

[0082] The program to be executed in the information processing device of this embodiment is provided pre-installed in a ROM 52 or the like.

[0083] The program executed by the information processing device of this embodiment may be configured to be provided as a computer program product by recording it in an installable or executable file format onto a computer-readable recording medium such as a CD-ROM (Compact Disk Read Only Memory), a flexible disk (FD), a CD-R (Compact Disk Recordable), or a DVD (Digital Versatile Disk).

[0084] Furthermore, the program executed by the information processing device of the embodiment may be stored on a computer connected to a network such as the Internet and provided by downloading it via the network. Alternatively, the program executed by the information processing device of the embodiment may be provided or distributed via a network such as the Internet.

[0085] The program executed in the information processing device of this embodiment can cause the computer to function as a component of the information processing device described above. This computer can read the program from a computer-readable storage medium onto the main memory and execute it using the CPU 51.

[0086] An example of the configuration of the embodiment is described below. (Configuration Example 1) A predictive model is constructed that takes multiple first time series data, each consisting of time series data for a first period of multiple input variables, as input and predicts one or more output variables at a time after the first period. For each of the one or more time points included in the first period, a third time series data is calculated, which is a time series data of an index representing the error or goodness of fit between one or more second time series data representing one or more of the output variables predicted by the prediction model, and the ground truth time series data representing the correct values ​​of one or more of the output variables at one or more time points included in the first period. Using the multiple first time series data and the third time series data, a time series causal graph is constructed that represents the causal relationship between the multiple input variables and the indicator. Processing unit, An information processing device equipped with the following features. (Configuration example 2) The multiple input variables include the actual value of the first variable and the predicted value of the first variable, The processing unit further uses the fourth time series data, which is time series data of the difference between the actual value and the predicted value, to construct the time series causal graph. The information processing device described in Configuration Example 1. (Configuration Example 3) The aforementioned processing unit, A combined time series data is generated by combining multiple first time series data and the third time series data. Using the aforementioned combined time series data, the time series causal graph is constructed. An information processing device as described in Configuration Example 1 or 2. (Configuration example 4) The aforementioned processing unit, Using the aforementioned time-series causal graph, a generative model is constructed that generates the input variable corresponding to the second node, which is a child node of the first node, from the input variable corresponding to the first node included in the time-series causal graph. Using the generative model described above, the contribution of multiple input variables to the index is calculated. The information processing device described in Configuration Example 3. (Configuration example 5) The generation model is a model that calculates a value for the variable corresponding to the second node by adding noise to the value calculated based on the input variable corresponding to the first node. The information processing device described in Configuration Example 4. (Configuration example 6) An information processing method performed by an information processing device, The steps include: constructing a predictive model that takes multiple first time series data, each of which is time series data for a first period of multiple input variables, as input and predicts one or more output variables at a time after the first period; A step of calculating a third time series data which is a time series data of an index representing the error or goodness of fit between one or more second time series data representing one or more output variables predicted by the prediction model for each of the one or more time points included in the first period, and the ground truth time series data representing the ground truth of one or more output variables at one or more time points included in the first period. The steps include constructing a time-series causal graph that represents the causal relationship between a plurality of input variables and an index, using a plurality of first time-series data and a third time-series data, Information processing methods including (Configuration example 7) On the computer, The steps include: constructing a predictive model that takes multiple first time series data, each of which is time series data for a first period of multiple input variables, as input and predicts one or more output variables at a time after the first period; A step of calculating a third time series data which is a time series data of an index representing the error or goodness of fit between one or more second time series data representing one or more output variables predicted by the prediction model for each of the one or more time points included in the first period, and the ground truth time series data representing the ground truth of one or more output variables at one or more time points included in the first period. The steps include constructing a time-series causal graph that represents the causal relationship between a plurality of input variables and an index, using a plurality of first time-series data and a third time-series data, A computer program designed to execute something.

[0087] While several embodiments of the present invention have been described, these embodiments are presented as examples only and are not intended to limit the scope of the invention. These novel embodiments can be carried out in a variety of other forms, and various omissions, substitutions, and modifications can be made without departing from the spirit of the invention. These embodiments and their variations are included in the scope and spirit of the invention, as well as in the claims of the invention and its equivalents. [Explanation of Symbols]

[0088] 100 Information Processing Devices 101 Acquisition Department 102 Output Control Unit 111 Predictive Model Construction Department 112 Indicator calculation section 113 Joint 114 Graph Construction Section 121 Generative Model Construction Section 122 Contribution Calculation Unit 131 Storage section

Claims

1. A predictive model is constructed that takes multiple first time series data, each consisting of time series data for a first period of multiple input variables, as input and predicts one or more output variables at a time after the first period. For each of the one or more time points included in the first period, a third time series data is calculated, which is a time series data of an index representing the error or goodness of fit between one or more second time series data representing one or more output variables predicted by the prediction model, and the correct time series data representing the correct values ​​of one or more output variables at one or more time points included in the first period. Using the multiple first time series data and the third time series data, a time series causal graph is constructed that represents the causal relationship between the multiple input variables and the indicator. Processing unit, An information processing device equipped with the following features.

2. The multiple input variables include the actual value of the first variable and the predicted value of the first variable. The processing unit further uses the fourth time series data, which is time series data of the difference between the actual value and the predicted value, to construct the time series causal graph. The information processing apparatus according to claim 1.

3. The aforementioned processing unit, Combined time series data is generated by combining multiple first time series data and third time series data. Using the aforementioned combined time series data, the time series causal graph is constructed. The information processing apparatus according to claim 1.

4. The aforementioned processing unit, Using the time-series causal graph, a generative model is constructed that generates the input variable corresponding to the second node, which is a child node of the first node, from the input variable corresponding to the first node included in the time-series causal graph. Using the generative model described above, the contribution of multiple input variables to the index is calculated. The information processing apparatus according to claim 3.

5. The generation model is a model that calculates a value by adding noise to a value calculated based on the input variable corresponding to the first node, and uses this value as the value of the variable corresponding to the second node. The information processing apparatus according to claim 4.

6. An information processing method performed by an information processing device, The steps include: constructing a predictive model that takes multiple first time series data, each of which is time series data for a first period of multiple input variables, as input and predicts one or more output variables at a time after the first period; A step of calculating a third time series data which is a time series data of an index representing the error or goodness of fit between one or more second time series data representing one or more output variables predicted by the prediction model for each of the one or more time points included in the first period, and ground truth time series data representing the correct values ​​of one or more output variables at one or more time points included in the first period. The steps include constructing a time-series causal graph that represents the causal relationship between a plurality of input variables and an index, using a plurality of first time-series data and a third time-series data, Information processing methods including

7. On the computer, The steps include: constructing a predictive model that takes multiple first time series data, each of which is time series data for a first period of multiple input variables, as input and predicts one or more output variables at a time after the first period; A step of calculating a third time series data which is a time series data of an index representing the error or goodness of fit between one or more second time series data representing one or more output variables predicted by the prediction model for each of the one or more time points included in the first period, and ground truth time series data representing the correct values ​​of one or more output variables at one or more time points included in the first period. The steps include constructing a time-series causal graph that represents the causal relationship between a plurality of input variables and an index, using a plurality of first time-series data and a third time-series data, A computer program designed to execute something.

Images