Predictive model evaluation apparatus, method, and program

Abstract

A prediction model evaluation device according to one embodiment of the present invention receives input of a new condition for a search for a causal relationship between a variable to be predicted and an explanatory variable that significantly correlates with the variable to be predicted and a new condition for generation of a prediction model, searches, on the basis of the new condition for the search, for a new causal relationship between the prediction variable and the explanatory variable that significantly correlates with the prediction variable, searches, on the basis of the newly retrieved causal relationship, for a new explanatory variable that has a parent relationship to the prediction variable, generates, on the basis of the new condition for the generation, a new prediction model that predicts the new explanatory variable that has the parent relationship to the prediction variable, and uses the newly generated prediction model to newly calculate a predicted value.

Description

[Technical Field]

[0001] Embodiments of the present invention relate to a predictive model evaluation apparatus, method, and program. [Background technology]

[0002] In areas related to a person's physical, economic, or social health (lifespan), such as healthcare, finance, or education, it is crucial for systems that intervene in individual decision-making to explain the circumstances or reasons for the intervention to the user and ensure they understand and accept the decision.

[0003] To achieve this, it is desirable that the system has the ability to accurately predict user behavior in advance from given information in order to understand the target user and their actions. In this regard, much research has focused on building highly accurate prediction models and developing techniques to identify features with high predictive power (see, for example, Non-Patent Documents 1 and 2). Here, high predictive power refers to the degree of contribution to minimizing prediction errors. In other words, high predictive power does not necessarily imply a causal relationship.

[0004] Furthermore, it is desirable that the system not only minimizes prediction errors but also understands the causes of the results. Conventional techniques have ensured the generalizability and robustness of prediction models by identifying causal relationships between data using statistical causal exploration methods and selecting variables based on these (see, for example, Non-Patent Document 3) (see, for example, Non-Patent Documents 4 and 5). [Prior art documents] [Non-patent literature]

[0005] [Non-Patent Document 1] Fisher, Aaron, Cynthia Rudin, and Francesca Dominici. "All Models are Wrong, but Many are Useful: Learning a Variable's Importance by Studying an Entire Class of Prediction Models Simultaneously." J. Mach. Learn. Res. 20.177 (2019): 1-81. [Non-Patent Document 2] Lundberg, Scott M., and Su-In Lee. "A unified approach to interpreting model predictions." Advances in neural information processing systems 30 (2017). [Non-Patent Document 3] Yu, Kui, Lin Liu, and Jiuyong Li. "A unified view of causal and non-causal feature selection." ACM Transactions on Knowledge Discovery from Data (TKDD) 15.4 (2021): 1-46. [Non-Patent Document 4] Li, Jundong, et al. "Feature selection: A data perspective." ACM computing surveys (CSUR) 50.6 (2017): 1-45. [Non-Patent Document 5] Susan Athey. 2017. Beyond prediction: Using big data for policy problems. Science 355, 6324 (2017), 483-485. [Overview of the project] [Problems that the invention aims to solve]

[0006] Conventional techniques can improve the predictive power of a machine learning model by identifying a group of other variables that are directly causal to a given variable in a given dataset and using these as predictor variables, thereby statistically supporting the causal relationship between these predictor variables and the system. However, they still suffer from the following problems.

[0007] Firstly, while statistical causal exploration methods guarantee causal relationships between variables from a statistical standpoint, they do not evaluate whether these causal relationships are understandable to humans. Specifically, the identified causal relationships and the structure and function of the model must be evaluated by a person who operates a machine learning model composed of features selected based on causal relationships identified by statistical causal exploration methods (hereinafter referred to as the "operator") and a person who receives information from this model (hereinafter referred to as the "explained person"). This is because the interpretability of a model is determined by whether the model operator or the person being explained can understand the model's behavior, and the more interpretable a model is, the more useful it is considered to be for individual users' decision-making.

[0008] Furthermore, due to the above-mentioned problems, it is not possible to understand the system settings that are favorable or unfavorable for the operator or the explained. It is expected that by using the operator or explained evaluation as feedback to the system, the system conditions can be updated to enable the presentation of information that is easier for the operator or the explained to interpret. However, currently, there is no evaluation from the operator or the explained for machine learning models composed of features selected by feature selection based on statistical causal search methods, or for systems in which this model is implemented. Therefore, this evaluation cannot be used to achieve more interpretable information presentation.

[0009] This invention was made in view of the above circumstances, and its purpose is to provide a predictive model evaluation device, method, and program that enable the construction of a predictive model for a variable to be predicted that is easy for the user to understand. [Means for solving the problem]

[0010] A predictive model evaluation device according to one aspect of the present invention includes: a search unit that searches for causal relationships between observed data of a variable to be predicted and observed data of explanatory variables that are significantly correlated with the variable to be predicted, and based on the searched causal relationships, finds explanatory variables that are parental to the variable to be predicted; a generation unit that generates a predictive model that predicts the variable to be predicted from the explanatory variables; a prediction unit that calculates a predictive value that is significantly correlated with the input data using the predictive model generated by the generation unit; and input of new conditions for the search for the causal relationships and new conditions for the generation of the predictive model by the generation unit. The system comprises an input unit for receiving data, the search unit newly searches for a causal relationship between the variable to be predicted and an explanatory variable that is significantly correlated with the variable to be predicted based on the new conditions of the search, and newly finds an explanatory variable that is a parent to the variable to be predicted based on the newly found causal relationship, the generation unit newly generates a prediction model that predicts the explanatory variable newly found by the search unit that is a parent to the variable to be predicted based on the new conditions of the generation, and the prediction unit newly calculates the predicted value using the prediction model newly generated by the generation unit.

[0011] A predictive model evaluation method according to one aspect of the present invention is a method performed by a predictive model evaluation device, which involves exploring the causal relationship between observed data of a variable to be predicted and observed data of an explanatory variable that is significantly correlated with the variable to be predicted; finding an explanatory variable that is related to the variable to be predicted based on the explored causal relationship; generating a predictive model that predicts the variable to be predicted from the explanatory variable; calculating a predictive value that is significantly correlated with the input data using the generated predictive model; accepting input of new conditions for exploring the causal relationship and new conditions for generating the predictive model; exploring a new causal relationship between the variable to be predicted and an explanatory variable that is significantly correlated with the variable to be predicted based on the new conditions for exploration; finding a new explanatory variable that is related to the variable to be predicted based on the newly explored causal relationship; generating a new predictive model that predicts the newly found explanatory variable that is related to the variable to be predicted based on the new conditions for generation; and calculating a new predictive value using the newly generated predictive model. [Effects of the Invention]

[0012] According to the present invention, it is possible to construct a predictive model for the variable to be predicted that is easy for users to understand. [Brief explanation of the drawing]

[0013] [Figure 1] Figure 1 shows an example of application of a predictive model evaluation device according to one embodiment of the present invention. [Figure 2A] Figure 2A shows an example of the information flow during the learning process. [Figure 2B] Figure 2B is a flowchart showing an example of the processing steps during learning. [Figure 3A] Figure 3A shows an example of the information flow during the explanation process. [Figure 3B] Figure 3B is a flowchart showing an example of the processing steps during the explanation. [Figure 3C]Figure 3C shows an example of the content to be presented during the explanation process. [Figure 4A] Figure 4A shows an example of the information flow during the evaluation process. [Figure 4B] Figure 4B is a flowchart showing an example of the processing steps during evaluation. [Figure 4C] Figure 4C is a flowchart showing an example of the processing steps during evaluation. [Figure 4D] Figure 4D shows an example of the content presented during the evaluation process. [Figure 5A] Figure 5A shows an example of the information flow during the update process. [Figure 5B] Figure 5B is a flowchart showing an example of the processing steps during an update. [Figure 6] Figure 6 is a block diagram showing an example of the hardware configuration of a predictive model evaluation device according to one embodiment of the present invention. [Modes for carrying out the invention]

[0014] Hereinafter, an embodiment of this invention will be described with reference to the drawings. In order to solve the problems of the conventional technology described above, this embodiment implements a machine learning model, which is a predictive model that identifies a direct causal system for a certain target variable based on the causal relationships between observed data extracted from given observational data using a statistical causal search method, and adopts this direct causal system as a feature. The system explains the causal relationships of the data and the behavior of the model, and accepts evaluations from the system operator or the person being explained to, and uses this as feedback, thereby providing a function that allows the user to interpret the causal relationships between the observed data. This enables the evaluation of causal feature prediction models by allowing humans to assess the interpretability of machine learning models and to understand the system design conditions that lead to information presentation that is easier for operators or those being explained to interpret.

[0015] Figure 1 shows an example configuration of a causal feature prediction model evaluation device according to one embodiment of the present invention. As shown in Figure 1, this device consists of five storage units and four mechanisms. Specifically, it comprises a data storage unit 100, a causal relationship storage unit 101, a learning model storage unit 102, an evaluation value storage unit 103, a system condition storage unit 104, a learning mechanism 200, an explanation mechanism 300, an evaluation mechanism 400, and an update mechanism 500. The learning mechanism 200 includes a preprocessing unit 201, a correlation detection unit 202, a causal search unit 203, and a learning unit (model generation unit) 204. The explanation mechanism 300 includes a prediction unit 301 and a presentation unit 302. The evaluation mechanism 400 includes an evaluation unit 401 and an adjustment unit 402. The update mechanism 500 includes a setting unit 501.

[0016] The data storage unit 100 stores each explanatory variable of the observed data and the target variable to be predicted (hereinafter sometimes referred to as the variable to be predicted). The causal relationship storage unit 101 stores the causal graph identified by the causal search unit 203. Here, the causal graph is a weighted directed graph, consisting of a set of nodes where each explanatory variable and the dependent variable are nodes, and an adjacency matrix relating the weights of the edges connecting the nodes.

[0017] The learning model storage unit 102 stores the model structure and learned parameters of the machine learning model learned by the learning unit 204. The evaluation value storage unit 103 stores evaluation values ​​for the system entered by the operator or the person being explained to (hereinafter collectively referred to as the user) in the evaluation unit 401. The items and measurement methods for the evaluation values ​​will be described later.

[0018] The system condition storage unit 104 stores the system conditions entered by the user in the adjustment unit 402 and the evaluation values ​​for the system entered in the evaluation unit 401, in an associated manner. Details of the system conditions will be described later.

[0019] Next, the processing of the causal feature prediction model evaluation device according to this embodiment will be described. This series of processes is divided into four stages: training, explanation, evaluation, and update.

[0020] <Process Overview> (1) Processing procedure during learning Figure 2A shows an example of the information flow during learning. Figure 2B is a flowchart showing an example of the learning process. The processing procedure follows the steps below. The specific procedures for these processes will be described later.

[0021] S100: The preprocessing unit 201 receives each explanatory variable of the observed data and the variable to be predicted from the data storage unit 100, and performs preprocessing such as (1) removal of missing values ​​and (2) normalization of each explanatory variable of the observed data, and outputs the variable to be predicted and the preprocessed explanatory variables to the correlation detection unit 202.

[0022] S101: The correlation detection unit 202 receives the variable to be predicted and the preprocessed explanatory variables output from the preprocessing unit 201 in S100, detects the preprocessed explanatory variables that are significantly correlated with the variable to be predicted, and outputs them to the causal search unit 203. At the same time, the correlation detection unit 202 also outputs the received variable to be predicted to the causal search unit 203.

[0023] S102: The causal exploration unit 203 receives the variable to be predicted and the pre-processed explanatory variables that are significantly correlated with this variable, output from the correlation detection unit 202 in S101, and generates a causal graph as the causal relationship between them using a statistical causal exploration method. The causal exploration unit 203 outputs this generated causal graph to the causal relationship storage unit 101, outputs the group of explanatory variables that have a parent relationship with the variable to be predicted in the causal graph and the variable to be predicted to the learning unit 204, and outputs the search conditions for causal exploration to the system condition storage unit 104.

[0024] S103: The learning unit 204 receives the explanatory variables that have a parent relationship with the variable to be predicted and the variable to be predicted, which were output in S102, from the causal search unit 203, constructs a machine learning model to predict the variable to be predicted from the explanatory variables, outputs the model structure, which is the structure of this machine learning model, and the trained parameters of the machine learning model to the learning model storage unit 102, and outputs the learning conditions of the machine learning model to the system condition storage unit 104.

[0025] (2) Processing procedure during explanation Figure 3A shows an example of the information flow during the explanation process. Figure 3B is a flowchart showing an example of the content of the explanation process. Figure 3C shows an example of the content to be presented during the explanation process. The processing procedure follows the steps below. The specific procedures for these processes will be described later.

[0026] S200: The data storage unit 100 receives input data from the operator of this device and stores it.

[0027] S201: The prediction unit 301 receives input data from the operator of this device, receives the trained model structure and parameters from the learning model storage unit 102, calculates a predicted value for the received input data, and outputs the input data, model structure, and predicted value to the presentation unit 302.

[0028] S202: The presentation unit 302 receives the causal graph from the causal relationship storage unit 101, the input data, model structure, and predicted values ​​output in S201 from the prediction unit 301, and the search conditions for causal exploration and the learning conditions for machine learning from the system condition storage unit 104. The presentation unit 302 then presents the received (1) causal graph (symbol a in Figure 3C), (2) predicted values ​​(symbol b in Figure 3C), (3) input data (symbol c in Figure 3C), (4) search conditions (drop-down selection type) (symbol d in Figure 3C), and (5) model structure and learning conditions (drop-down selection type) (symbol e in Figure 3C) to the operator of the device, for example, as a presentation screen G1 shown in Figure 3C.

[0029] (3) Processing procedures during evaluation Figure 4A shows an example of the information flow during the evaluation process. Figures 4B and 4C are flowcharts showing an example of the content of the evaluation process. Figure 4D shows an example of the content presented during the evaluation process. The processing procedure follows the steps below. The specific procedures for these processes will be described later.

[0030] S300: The evaluation unit 401 receives input of evaluation values ​​for the explanation of the prediction results of the prediction model from the operator of this device and outputs these to the evaluation value storage unit 103.

[0031] S301: The evaluation value storage unit 103 receives the evaluation value output in S300 from the evaluation unit 401 and outputs it to the system condition storage unit 104.

[0032] S302: The system condition storage unit 104 receives the evaluation value output in S301 from the evaluation value storage unit 103, matches this evaluation value with the system condition to be evaluated, and stores it.

[0033] S303: The adjustment unit 402 receives input of the model structure, search conditions, and learning conditions of the prediction model from the operator of this device, outputs the search conditions and learning conditions to the system condition storage unit 104, outputs the search conditions to the causal search unit 203, and outputs the model structure and learning conditions to the learning unit 204.

[0034] S304: The causal search unit 203 receives the search conditions output in S303 from the adjustment unit 402, generates a causal graph according to these search conditions, outputs the generated causal graph to the causal relationship storage unit 101, and outputs the group of explanatory variables that have a parent relationship with the variable to be predicted in the generated causal graph, along with the variable to be predicted, to the learning unit 204.

[0035] S305: The learning unit 204 receives the group of explanatory variables that have a parent relationship with the variable to be predicted and the variable to be predicted, which are output from the causal search unit 203 in S304, and the learning conditions output from the adjustment unit 402 in S303. It constructs a machine learning model to predict the variable to be predicted from the group of explanatory variables, and outputs this learning model, model structure, and parameters to the learning model storage unit 102.

[0036] S306: The prediction unit 301 receives the input data stored during the explanation from the data storage unit 100, receives the trained model structure and parameters from the trained model storage unit 102, calculates the predicted value for the input data, and outputs the input data, model structure, and predicted value to the presentation unit 302.

[0037] S307: The presentation unit 302 receives a causal graph from the causal relationship storage unit 101 and input data, model structure, and predicted values ​​from the prediction unit 301, and presents the prediction results and their explanation to the operator of the device. During evaluation, the presentation screen G2 shown in Figure 4D displays the interface in the evaluation unit 401 (symbol a in Figure 4D) and the interface in the adjustment unit 402 (symbol b in Figure 4D).

[0038] (4) Processing procedure during updates Figure 5A shows an example of the information flow during the update process. Figure 5B is a flowchart showing an example of the update process. The processing procedure follows the steps below. The specific procedures for these processes will be described later.

[0039] S400: The setting unit 501 receives system design conditions and evaluation values ​​from the system condition storage unit 104, calculates the optimal combination of search conditions, model structure, and learning conditions for the operator of this device, outputs the search conditions to the causal search unit 203, and outputs the model structure and learning conditions to the learning unit 204.

[0040] S401: The causal exploration unit 203 receives the exploration conditions that are optimal for the operator of this device and output in S400 from the setting unit 501, and sets them as the exploration conditions of the causal exploration unit 203.

[0041] S402: The learning unit 204 receives the model structure and learning conditions that are optimal for the operator and output in S400 from the setting unit 501, and sets them as the model structure and exploration conditions of the learning unit 204.

[0042] <Details of the process> The specific procedure of each processing step is described below. Hereinafter, the number of samples of the target data is denoted as n, and the number of explanatory variables is denoted as m. Also, for a certain sample i ∈ {1,..., n} and a certain variable j ∈ {1,..., m}, the data value of variable j is denoted as the data value x of the explanatory variable ij Let it be. Also, for a certain sample i ∈ {1,..., n}, the data value of the target variable is denoted as y i Let it be. The data value x ij and the data value y i are stored in the data storage unit 100.

[0043] (1) Specific processing content during learning The procedure of the processing in the preprocessing unit 201 (S100 in FIG. 2B) is described. The preprocessing unit 201 receives the data value x of the explanatory variable and the data value y of the target variable from the data storage unit 100, and performs necessary preprocessing on the data value x and the data value y in order to execute the correlation analysis in the correlation detection unit 202 included in the learning mechanism 200 and the causal exploration included in the causal exploration unit 203. Here, first, missing values are excluded, and then the explanatory variables are normalized. ij and the data value y of the target variable i to perform necessary preprocessing on the data value x ij and the data value y i . Here, first, missing values are excluded, and then the explanatory variables are normalized.

[0044] Next, the above exclusion of missing values is described. In the exclusion of missing values, the preprocessing unit 201 constructs the data sample set D from the data value x and the data value y received from the data storage unit 100 according to the following formulas (1) and (2). ij and the data value y i according to the following formulas (1) and (2).

[0045] d i ={y i ,x i1 ,…,x im} …Equation (1) D={d i |i=1,…,n} …Equation (2) Here, the preprocessing unit 201 performs the following processing on all i according to equation (3), thereby generating a data sample d containing missing values. i Remove the missing values ​​from the data sample set D. Here, data values ​​indicating missing values ​​are denoted as NA.

[0046] if NA∈d i then D→D\{d i} …Equation (3) The data sample set obtained through the above procedure is D * Let's assume that. For simplicity, we will use D here. * Assuming that = D, the following procedures will be described.

[0047] Next, we will discuss the normalization of the explanatory variables mentioned above. In this normalization, the preprocessor 201 processes the data sample set D mentioned above. * The explanatory variable x ij (∈d i )∈D * Next, we normalize the variable j. Here, following the procedures in equations (4), (5-1), and (5-2), we normalize the data values ​​of the explanatory variables x such that the mean of the variable j is 0 and the variance is 1. ij The standardized variable z ij To obtain.

[0048]

number

[0049] The preprocessing unit 201 processes the data value y of the target variable. i ∈D * and the standardized variable z ij This is output to the correlation detection unit 202.

[0050] Next, the processing procedure in the correlation detection unit 202 (S101 in Figure 2B) will be described. The correlation detection unit 202 receives the data value y of the target variable from the preprocessing unit 201. i and the standardized variable z ij The target variable y=(y1,…,y n ) T The variable z is correlated with j =( z 1j ,…,z nj ) T Identify the variable z that has the above correlation with the target variable y. j Correlation coefficient γ i The result is obtained by the following equation (6).

[0051]

number

[0052] Here, Cov in equation (6) is the covariance of the two given variables, and Var in equation (6) is the function used to calculate the variance of the given variables. Letting α be the significance level of the correlation analysis, the significance probability of the correlation coefficient obtained in equation (6) is p. j When p j The set of variables Z satisfying <α * The significance level α is defined as shown in equation (7) below. Generally, α is set to 0.05, 0.01, or 0.001, but the operator of this device may set the significance level α arbitrarily. Z * ={z j |p j <α,j∈{1,…,m}} …Equation (7) For simplicity, here we'll use Z={z1…,z m When}, Z * Assuming that = Z, the following procedures will be described. The correlation detection unit 202 analyzes the target variable y and the set of variables Z. * The output is sent to the causal search unit 203.

[0053] Next, we will describe the processing procedure in the causal search unit 203 (S102 in Figure 2B). The causal search unit 203 receives the target variable y and the set of variables Z from the correlation detection unit 202. * We use this method to explore and estimate the causal relationships between these variables. Here, we use a method called LiNGAM (Linear Non-Gaussian Acyclic Model). However, any method that can explore causal relationships and evaluate their statistical reliability is acceptable for conducting causal exploration. LiNGAM is one of the statistical causal exploration methods, and by setting the following conditions (1) to (4), it uniquely identifies the causal relationships between variables.

[0054] (1) Conditions regarding the function system: The relationships between variables are linear. (2) Conditions regarding error distribution: The error distributions of the variables are non-Gaussian processes and are independent of each other. (3) Conditions regarding cyclicity: The causal relationships between variables are acyclic. (4) Conditions regarding unobserved common causes: No unobserved common causes exist. Under these conditions, the causal search unit 203 assumes a model using LiNGAM according to the following equation (8).

[0055]

number

[0056] In equation (8), each variable is v i =(y,z1,…,z m ) and the coefficient is b ij Let the error variable be e i Let's assume this. In matrix notation, it can be written as equation (9) below.

[0057]

number

[0058] In causal exploration using LiNGAM, the causal exploration unit 203 first calculates the value of each element of the coefficient matrix (where the order and scale are not distinguishable), then estimates the causal order between variables, and finally calculates the coefficient matrix.

[0059]

number

[0060] We estimate the causal order. A causal order is an order in which, when variables are rearranged according to that order, a later variable will never be the cause (ancestor in the graph representation) of a earlier variable. Here, the variable v i Let k(i) be the causal order of the variables. For example, if k(3)=2, then the causal order of variable v3 is interpreted as being second.

[0061] Next, we will discuss the calculation of the values ​​of each element of the coefficient matrix. From equation (9) above...

[0062]

number

[0063] This is obtained. However, the following relationship holds:

[0064]

number

[0065] Here,

[0066]

number

[0067] Since each element is independent of the differences, independent component analysis can be performed.

[0068]

number

[0069] The inverse matrix

[0070]

number

[0071] The following can be obtained as the restoration matrix. The coefficient matrix estimated by this procedure is

[0072]

number

[0073] as,

[0074]

number

[0075] This is obtained. However, since this value is obtained by independent component analysis, the order and scale of the matrix are not distinguishable. Therefore, the causal search unit 203 corrects the order of the matrix by estimating the causal order as described below.

[0076] Next, we will describe the method for estimating the causal order. First, we add a permutation matrix to both sides of equation (9) from left to right.

[0077]

number

[0078] Multiplying by gives the following equation (10).

[0079]

number

[0080] Considering the property that a permutation matrix is ​​an orthogonal matrix,

[0081]

number

[0082] Since this holds true, if we introduce this into the first term on the right-hand side of equation (10), we obtain the following equation (11).

[0083]

number

[0084] This is a permutation matrix.

[0085]

number

[0086] Variable vector with the order of elements swapped by

[0087]

number

[0088] This can be considered as a LiNGAM model. In this case, the coefficient matrix is

[0089]

number

[0090] In the LiNGAM model, the coefficient matrix follows a causal order.

[0091]

number

[0092] When this is constructed, the coefficient matrix

[0093]

number

[0094] It has the property of being close to a strictly lower triangular matrix. Using this, the causal search unit 203,

[0095]

number

[0096] Such that the matrix is ​​a lower triangular matrix

[0097]

number

[0098] Search for it.

[0099]

number

[0100] Whether it is close to an exact lower triangular matrix can be measured by the sum of the squares of the upper triangular and diagonal elements. Therefore, the causal search unit 203 solves the optimization problem shown in equation (12) below,

[0101]

number

[0102] Such that the matrix is ​​a lower triangular matrix

[0103]

number

[0104] The causal search unit 203 calculates the following:

[0105]

number

[0106] of

[0107]

number

[0108] Replace it with this.

[0109]

number

[0110] As a result, the causal search unit 203 obtains a causal order.

[0111]

number

[0112] This refers to k(i) = j.

[0113] Next, we will describe the method for estimating the coefficient matrix described above. First, we consider the model shown in equation (13) below, according to the causal order obtained above.

[0114]

number

[0115] This is a variable v i This is a model that regresses on all variables with a smaller causal order than the given one. In other words, we just need to find the partial regression coefficients when the above equation (13) is treated as a linear regression model. Here, we use a type of sparse regression called adaptive Lasso. Specifically, the causal search unit 203 obtains the partial regression coefficients by solving the optimization problem shown in the following equation (14).

[0116]

number

[0117] Here, λ and γ in equation (14) represent the adjustment parameters, and in equation (14)

[0118]

number

[0119] is, b ij The consistent estimator, estimated using linear regression by least squares, is shown. λ and γ are arbitrarily determined by the operator of this device.

[0120] By applying the bootstrap method to the above procedure, the causal search unit 203 evaluates the statistical reliability of the obtained coefficient matrix. Specifically, the causal search unit 203 evaluates the statistical reliability of the coefficient matrix for N sample extractions. i The probability that ≠ 0 is calculated. The causal search unit 203 uses this probability value q to calculate the probability. ij A matrix with (i,j) as its components

[0121]

number

[0122] Let's assume that.

[0123] The causal search unit 203 identifies explanatory variables z whose bootstrap probability is greater than or equal to c for the target variable y (=v0). j (=v1,…,v m ) is obtained by the procedure shown in equation (15) below.

[0124] Z c ={z j |q 0j >c,j=1,…,m} …Equation (15) The reference probability c in equation (15) can be arbitrarily determined by the operator of this device.

[0125] The causal search unit 203 assigns a node set V={v0,v1,…,v} to the causal relationship storage unit 101. m Let} be the adjacency matrix between nodes, and the coefficient matrix be the adjacency matrix between nodes.

[0126]

number

[0127] A causal graph

[0128]

number

[0129] and bootstrap probability matrix

[0130]

number

[0131] The output is also generated. The causal search unit 203 also searches for conditions related to the function system of the causal search method, conditions related to the error distribution, conditions related to cyclicity, conditions related to unobserved common causes, adjustment parameters λ and γ, and reference probability c as search conditions C. D The search condition C is stored in the system condition storage unit 104. D The output is generated, and the learning unit 204 receives the target variable y and the group of explanatory variables Z. c Outputs.

[0132] Next, we will describe the processing procedure in the learning unit 204 (S103 in Figure 2B). The learning unit 204 receives the target variable y and the group of explanatory variables Z from the causal search unit 203. c Receiving the explanatory variables Z c We calculate the parameters for predicting the target variable y from this data.

[0133] Here, the model structure to be adopted is

[0134]

number

[0135] Let's assume that here we have a linear regression model.

[0136]

number

[0137] It is used as follows. However, the explanatory variable group Z c If it is possible to calculate the parameters for predicting the dependent variable y from the data, the model is not limited to a linear regression model.

[0138] First, consider the model shown in equation (16) below.

[0139]

number

[0140] Here,

[0141]

number

[0142] Therefore, the following equation (17) holds true. y=X T θ+e...Equation (17) And the optimization method used for parameter estimation is

[0143]

number

[0144] Let's assume that here we use the least squares method.

[0145]

number

[0146] It is used as follows. However, the parameter estimation method is not limited to the least squares method if the optimal parameters can be obtained. By applying the least squares method, the optimal parameter Θ shown in equation (18) below is obtained.

[0147] Θ=(XX T ) -1 Xy...Formula (18) The learning unit 204 stores the model structure in the learning model storage unit 102.

[0148]

number

[0149] The learned parameters Θ are output. The learning unit 204 also outputs the model structure

[0150]

number

[0151] and explanatory variable Z c and parameter optimization methods

[0152]

number

[0153] Learning condition C L The system condition storage unit 104 stores the learning condition C L Outputs.

[0154] (2) Specific details of the process to be explained Next, the processing procedure in the data storage unit 100 during the explanation will be described (S200 in Figure 3B). The data storage unit 100 receives new data input from the operator of this device and stores it.

[0155] Next, the processing procedure in the prediction unit 301 will be described (S201 in Figure 3B). The prediction unit 301 receives the model structure from the learning model storage unit 102.

[0156]

number

[0157] Based on the learned parameters Θ, the predicted value of the input data of the operating entity of this device

[0158]

Number

[0159] is calculated. The prediction unit 301 uses the input data as x input and, as shown in the following formula (19)

[0160]

Number

[0161] is obtained.

[0162]

Number

[0163] The prediction unit 301 sends the input data x input , the model structure

[0164]

Number

[0165] , and the predicted value

[0166]

Number

[0167] to the presentation unit 302 for output.

[0168] Next, the procedure of the processing in the presentation unit 302 will be described (S202 in FIG. 3B). The presentation unit 302 receives the predicted value from the prediction unit 301

[0169]

Number

[0170] are presented to the operator of this device as prediction results. Further, the presentation unit 302 presents the input data x received from the prediction unit 301 input to the operator of this device as the history data of the operator of this device. Further, the presentation unit 302 receives the search condition C D and the learning condition C L from the system condition storage unit 104, and presents these to the operator of this device as auxiliary information for explaining that prediction values have been obtained. Further, the presentation unit 302 receives the causal graph

[0171] [Number]

[0172] and the bootstrap probability matrix

[0173] [Number]

[0174] and presents to the operator of this device a subgraph G'⊂G of the causal graph G consisting of the node v∈{y}∩Z c and the coefficient b between the node v ij .

[0175] Examples of presenting some of this information are shown at the upper part of the interface example shown in FIG. 3C. Here, it is assumed that when input data (reference symbol c in FIG. 3C) regarding lifestyle during a diet period in the past one month is given, the rebound rate in the subsequent one month is predicted, and the prediction result is explained to the operator of this device by the screen G1. The example shown in Figure 3C assumes that the causal search method and machine learning method, listed in the columns for information on prediction methods (indicated by labels d and e in Figure 3C), will be displayed in a dropdown format so that the user operating the device can select them. When the update button on the screen is selected, (1) the causal graph (symbol a in Figure 3C) and (2) the predicted value (symbol b in Figure 3C) are updated according to the conditions selected by the operator of this device via the dropdown menu. Details of this procedure will be described later in the processing procedure in the adjustment unit 402 (S303 in Figure 4B).

[0176] (3) Specific details of the process during evaluation Next, the processing procedure in the evaluation unit 401 will be described (S300 in Figure 4B). The evaluation unit 401 receives the prediction result from the device and an evaluation value regarding the interpretability of its explanation as input from the operator of the device.

[0177] Here, the evaluation items for interpretability are subjective evaluations of the information presented by the system from the perspective of interpretability, such as whether the operator of the system finds the prediction results, causes, the process by which they were calculated, the behavior of the model, and other information accompanying the explanation of the prediction results easy to understand, convincing, and reliable. The question and answer formats are arbitrary as long as they measure this evaluation. Here, as an example, the interface of the evaluation unit 401 (labeled a in Figure 4D), as shown in the interface example in Figure 4D, asks for the degree of understanding of the cause of the prediction result on a three-point scale, and accepts the answer as an evaluation value for the interpretability of the operator of this device.

[0178] The evaluation unit 401 receives an evaluation set of interpretability values ​​input by the operator of this device, and this set of values ​​is denoted as R. The evaluation unit 401 outputs the evaluation value set R to the evaluation value storage unit 103.

[0179] Next, the procedure of the process in the evaluation value storage unit 103 will be described (S301 in FIG. 4B). The evaluation value storage unit 103 receives the set of evaluation values R output from the evaluation unit 401 in S300 and stores it in the evaluation value storage unit 103. Also, the evaluation value storage unit 103 receives the set of evaluation values R output from the evaluation unit 401 in S300 and outputs it to the system condition storage unit 104.

[0180] Next, the procedure of the process in the system condition storage unit 104 will be described (S302 in FIG. 4B). The system condition storage unit 104 receives the set of evaluation values R from the evaluation value storage unit 103, and the search condition C received from the causal search unit 203 during learning D and the learning condition C received from the learning unit 204 during learning L and associates them as <R, C D , C L > and stores this.

[0181] Next, the procedure of the process in the adjustment unit 402 will be described (S303 in FIG. 4B). The adjustment unit 402 aims to present a new prediction result after the operator of this device adjusts the prediction result by this device and the search condition C D and the learning condition C L (models and parameters of the causal search method and the machine learning method). It accepts new models and parameters as input from the operator of this device on the interface (reference symbol b in FIG. 4D), and uses this as a new search condition

[0182]

Number

[0183] ​​​​​​​​​​​​​​​​

number

[0187] The learning conditions are set in the learning unit 204.

[0188]

number

[0189] The output is as follows: Here, as an example, we will describe the case where the baseline probability c = 0.50 → 0.30 adopted by the causal search unit 203 during training is updated. Assume that there are no changes to the training conditions.

[0190] Next, the processing in the causal search unit 203 (S304 in Figure 4B) will be explained. This processing is performed by the causal search unit 203 when the search conditions are met.

[0191]

number

[0192] After being updated, the same processing as during training (S102) is performed. Here, the search conditions are

[0193]

number

[0194] As a result of the update, that is, the reference probability c was updated from 0.50 to 0.30, the explanatory variables

[0195]

number

[0196] The system condition storage unit 104 stores the search conditions.

[0197]

number

[0198] It outputs the following to the learning unit 204

[0199]

number

[0200] Outputs.

[0201] Next, we will explain the processing in the learning unit 204 (S305 in Figure 4B). This processing involves the learning unit 204 updating the explanatory variables from the causal search unit 203.

[0202]

number

[0203] Upon receiving, the explanatory variables

[0204]

number

[0205] After being changed, the process becomes the same as during training (S103). Consequently, the trained parameters are

[0206]

number

[0207] The learning unit 204 then updates the learning parameters.

[0208]

number

[0209] This is output to the learning model storage unit 102.

[0210] Next, the processing in the prediction unit 301 (Figure 4CS306) will be explained. This processing involves the prediction unit 301 receiving the updated learning parameters from the learning unit 204.

[0211]

number

[0212] It receives the parameter and

[0213]

number

[0214] After being changed, the process becomes the same as during the explanation (S201). Consequently, the predicted value

[0215]

number

[0216] The prediction unit 301 then updates the updated prediction value.

[0217]

number

[0218] The output is sent to the display unit 302.

[0219] Next, we will describe the processing in the presentation unit 302 (S307 in Figure 4C). In this process, the presentation unit 302 receives the updated predicted value from the prediction unit 301.

[0220]

number

[0221] Receive the predicted value

[0222] [Number]

[0223] is changed to, and the predicted value after update

[0224] [Number]

[0225] is presented to the operator of this device. Also, the presentation unit 302 obtains a causal graph

[0226] [Number]

[0227] and a bootstrap probability matrix

[0228] [Number]

[0229] receives, and presents a sub - graph G´⊂G of the causal graph G, which consists of the coefficient b

[0230] [Number]

[0231] between node ij and node v, to the operator of this device.

[0232] (4) Content of specific processing during update Next, the processing in the setting unit 501 during update will be described (S400 in Fig. 5B). The setting unit 501 is activated at an arbitrary timing according to the command of the operator of this device. For example, methods such as explicitly commanding, or commanding to be activated regularly can be cited. The processing by the setting unit 501 aims to set, for the causal search unit 203 and the learning unit 204, the search conditions and learning conditions that are estimated to have the highest interpretability for the operator of this device, using the information on the search conditions, learning conditions, and evaluation value sets stored in the system condition storage unit 104. In the system condition storage unit 104, there is a certain search condition C D and a certain learning condition C L and a certain set of evaluation values R corresponding to these, <R, C D , C L > which is stored as information in one unit. Hereinafter, the <R, C D , C L > recorded at the i-th position is regarded as the condition evaluation unit

[0233]

Number

[0234] (the i-th one is earlier than the (i + 1)-th one).

[0235] The setting unit 501 first receives the condition evaluation unit u i from the system condition storage unit 104. Here, the setting unit 501 may receive all the condition evaluation units stored in the system condition storage unit 104. Also, the setting unit 501 may receive a part of all the condition evaluation units stored in the system condition storage unit 104. For example, methods such as the setting unit 501 receiving the condition evaluation units recorded during the most recent one week, or receiving the condition evaluation units randomly extracted 50% of the whole can be mentioned. Here, assuming that the setting unit 501 has received all the condition evaluation units, the subsequent processing will be described.

[0236] The setting unit 501 selects, from all the condition evaluation units u i received from the system condition storage unit 104, the following condition evaluation unit with the highest interpretability evaluation value and the most recently recorded

[0237]

number

[0238] Extract it.

[0239]

number

[0240] The setting unit 501 is connected to the causal search unit 203.

[0241]

number

[0242] It outputs the following to the learning unit 204

[0243]

number

[0244] Outputs.

[0245] In the above, the extraction criteria are the highest interpretability score and the most recently recorded data. However, the method of extracting condition evaluation units is not limited to the above, as long as the objective of setting the causal search unit 203 and the learning unit 204 with the search and learning conditions that are estimated to be the most interpretable for the operator of this device by the setting unit 501 is achieved. For example, one method is to select the search and learning conditions that have the highest interpretability score and are also the most frequently used among them.

[0246] Next, we will describe the processing in the causal search unit 203 (S401 in Figure 5B). The causal search unit 203 receives from the setting unit 501.

[0247]

number

[0248] Upon receiving the information, the causal search conditions of the causal search unit 203 are set.

[0249]

number

[0250] Update it.

[0251] Next, we will describe the processing in the learning unit 204 during the update (Figure 5BS402). The learning unit 204 receives from the setting unit 501.

[0252]

number

[0253] Upon receiving, the learning conditions of learning unit 204

[0254]

number

[0255] Update it.

[0256] (Effects of this embodiment) In this embodiment, by utilizing a statistical causal search method, it is possible to evaluate whether the causal relationships between variables obtained are understandable to the human operator of this device by associating them with the conditions of the model. In this embodiment, the system settings that are favorable or unfavorable to the operator of this device can be identified. Based on the two points mentioned above, in this embodiment, the interpretability of information presented by the machine learning model can be improved from a human perspective, making it possible to construct a predictive model of the target variable that is easy for users to understand.

[0257] Figure 6 is a block diagram showing an example of the hardware configuration of a predictive model evaluation device according to one embodiment of the present invention. In the example shown in Figure 6, the prediction model evaluation device 10 according to the above embodiment is composed of, for example, a server computer or a personal computer, and has a hardware processor 11A such as a CPU. A program memory 11B, a data memory 12, an input / output interface 13, and a communication interface 14 are connected to this hardware processor 11A via a bus 15.

[0258] The communication interface 14 includes, for example, one or more wireless communication interface units, enabling the transmission and reception of information with the communication network NW. As the wireless interface, for example, an interface employing a low-power wireless data communication standard such as a wireless LAN (Local Area Network) is used.

[0259] The input / output interface 13 is connected to an input device 60 and an output device 70 that are attached to the predictive model evaluation device 10 and used by users or the like. The input / output interface 13 receives operation data input by a user or the like through an input device 60 such as a keyboard, touch panel, touchpad, or mouse, and outputs output data to an output device 70, including a display device using liquid crystal or organic EL (electroluminescence), for display. The input device 60 and output device 70 may be devices built into the predictive model evaluation device 10, or they may be input and output devices of other information terminals that can communicate with the predictive model evaluation device 10 via a network NW.

[0260] The program memory 11B is a non-temporary tangible storage medium in which a non-volatile memory that can be written to and read at any time, such as an HDD (Hard Disk Drive) or SSD (Solid State Drive), is used in combination with another non-volatile memory such as ROM (Read Only Memory), and stores programs necessary for executing various control processes according to one embodiment.

[0261] The data memory 12 is a tangible storage medium that, for example, uses a combination of the above-mentioned non-volatile memory and volatile memory such as RAM (Random Access Memory), and is used to store various data acquired and created during the process of various operations.

[0262] A predictive model evaluation device 10 according to one embodiment of the present invention may be configured as a data processing device having a software-based processing function unit. The various storage units used as work memory and other functions by the predictive model evaluation device 10 can be configured using the data memory 12 shown in Figure 6. However, these storage areas are not essential components within the predictive model evaluation device 10; for example, they may be external storage media such as USB (Universal Serial Bus) memory, or areas provided in storage devices such as database servers located in the cloud.

[0263] The processing function described above can be implemented by having the hardware processor 11A read and execute a program stored in the program memory 11B. This processing function may also be implemented in various other forms, including integrated circuits such as Application Specific Integrated Circuits (ASICs) or Field-Programmable Gate Arrays (FPGAs).

[0264] Furthermore, the methods described in each embodiment can be stored as programs (software means) that can be executed by a computer, such as magnetic disks (floppy disks, hard disks, etc.), optical disks (CD-ROMs, DVDs, MOs, etc.), and semiconductor memories (ROMs, RAMs, flash memory, etc.), and can also be transmitted and distributed via communication media. The programs stored on the media also include configuration programs that configure the computer with software means (including not only the execution program but also tables and data structures) to be executed by the computer. The computer implementing this device reads the program recorded on the recording medium and, if necessary, constructs the software means using the configuration program, and executes the above-described processes by controlling the operation of this software means. The recording medium referred to in this specification is not limited to distribution media, but also includes storage media such as magnetic disks and semiconductor memories provided inside the computer or in devices connected via a network.

[0265] It should be noted that the present invention is not limited to the embodiments described above, and can be modified in various ways during implementation without departing from its essence. Furthermore, each embodiment may be combined as appropriate, and in that case, the combined effects can be obtained. Moreover, the above embodiments include various inventions, and various inventions can be extracted by selecting combinations from the multiple constituent elements disclosed. For example, if the problem can be solved and effects obtained even if some constituent elements are deleted from all the constituent elements shown in the embodiment, then the configuration with these deleted constituent elements can be extracted as an invention. [Explanation of symbols]

[0266] 10…Predictive model evaluation device 100...Data storage unit 101...Causal Relationship Storage Unit 102...Learning model storage unit 103...Evaluation value storage unit 104... System condition storage unit 200...Learning mechanism 201…Pre-processing section 202...Correlation detection unit 203…Cause and effect search department 204…Learning Department 300...Explanatory mechanism 301... Prediction Department 302…Presentation part 400…Evaluation Organization 401…Evaluation Department 402...Adjustment section 500…Update mechanism 501...Settings Department

Claims

1. A search unit that searches for causal relationships between observed data of a variable to be predicted and observed data of explanatory variables that are significantly correlated with the variable to be predicted, and based on the searched causal relationships, finds explanatory variables that are related to the variable to be predicted as parents. A generation unit that generates a prediction model that predicts the variable to be predicted from the explanatory variables, A prediction unit that calculates a predicted value that is significantly correlated with the input data using the prediction model generated by the generation unit, An input unit that accepts input of new conditions for exploring the causal relationship and new conditions for generating the prediction model by the generation unit, Equipped with, The search unit, Based on the new conditions of the aforementioned search, a new causal relationship is explored between the variable to be predicted and explanatory variables that are significantly correlated with the variable to be predicted. Based on the newly discovered causal relationships, we newly determine the explanatory variables that have a parent relationship with the variable to be predicted. The generating unit is Based on the new generation conditions, a new predictive model is generated that predicts the explanatory variables newly obtained by the search unit that have a parent relationship with the variable to be predicted. The prediction unit, The prediction model newly generated by the generation unit is used to newly calculate the prediction value. Predictive model evaluation device.

2. The system further includes a preprocessing unit that performs preprocessing to normalize the explanatory variables by excluding missing values ​​in the observed data of the prediction target and the observed data of the explanatory variables. The search unit, The causal relationship between the observed data of the variable to be predicted and the observed data of the explanatory variable that is significantly correlated with the variable to be predicted and has been processed by the preprocessing unit is explored. The predictive model evaluation device according to claim 1.

3. The system further comprises the conditions for the search unit to search for the causal relationship, the conditions for the generation unit to generate the prediction model, and an output unit that outputs the causal relationship searched by the search unit. The aforementioned input unit is The system further accepts input of evaluation results for the predicted values ​​predicted by the prediction unit and evaluation results for the results output by the output unit. The predictive model evaluation device according to claim 1.

4. The prediction unit, Calculate a predicted value for each of the multiple input data, The aforementioned input unit is The prediction unit receives input of evaluation results for each of the predicted values ​​predicted by the unit. The system further includes a setting unit that sets the structure of the prediction model when the evaluation result is optimal, and the conditions output by the output unit when the evaluation result is optimal, based on the evaluation result and the conditions output by the output unit. The predictive model evaluation device according to claim 3.

5. A method performed by a predictive model evaluation device, The causal relationship between the observed data of the variable to be predicted and the observed data of explanatory variables that are significantly correlated with the variable to be predicted is explored, and based on the explored causal relationship, the explanatory variables that have a parent relationship with the variable to be predicted are determined. A predictive model is generated that predicts the variable to be predicted from the explanatory variables. Using the generated prediction model, we calculate predicted values ​​that are significantly correlated with the input data. The system accepts input of new conditions for exploring the aforementioned causal relationship and new conditions for generating the aforementioned predictive model. Based on the new conditions of the aforementioned search, a new causal relationship is explored between the variable to be predicted and explanatory variables that are significantly correlated with the variable to be predicted. Based on the newly discovered causal relationships, we newly determine the explanatory variables that have a parent relationship with the variable to be predicted. Based on the new generation conditions, a new predictive model is generated that predicts the newly obtained explanatory variables that have a parent relationship with the variable to be predicted. Using the newly generated prediction model, the predicted values ​​are newly calculated. Methods for evaluating predictive models.

6. A predictive model evaluation processing program that causes a processor to function as one of the components of the predictive model evaluation apparatus according to any one of claims 1 to 4.