Information processing device, information processing method, and program

The information processing device uses multiple subsets of feature quantities as explanatory variables to enhance abnormal case detection accuracy by accounting for interdependent feature values, improving the identification of cases where multiple features are abnormal.

JP7586342B2Active Publication Date: 2024-11-19NEC CORP
View PDF 3 Cites 0 Cited by

Patent Information

Application Number
JP2023556013
Authority / Receiving Office
JP · JP
Patent Type
Patents
Current Assignee / Owner
Filing Date
2021-10-29
Publication Date
2024-11-19
Estimated Expiration
2041-10-29

AI Technical Summary

Technical Problem

Existing methods fail to accurately detect abnormal cases when multiple feature amounts become abnormal values simultaneously, as they do not account for the interdependence of these features.

Method used

An information processing device that uses multiple different subsets of feature quantities as explanatory variables to output multiple prediction results, calculating an abnormality degree by referencing these results.

Benefits of technology

This approach allows for more accurate detection of abnormal cases by considering the interdependence of feature values, reducing the likelihood of misclassifying cases as normal when multiple features are abnormal.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 0007586342000006
    Figure 0007586342000006
  • Figure 0007586342000007
    Figure 0007586342000007
  • Figure 0007586342000008
    Figure 0007586342000008
Patent Text Reader

Abstract

To solve the problem of detecting an abnormal case more accurately, this information processing device comprises: an acquisition means (21) that acquires a case represented as a set of a plurality of features; a prediction means (22) that outputs a plurality of prediction results by setting at least one feature included in the case as an objective variable and setting, as explanatory variables, a plurality of subsets of features that are other than the at least one feature and are different from each other among the plurality of features; and an abnormality level output means (23) that outputs an abnormality level of the case by referring to the plurality of prediction results.
Need to check novelty before this filing date? Find Prior Art

Description

[Technical field]

[0001] The present invention relates to an information processing device, an information processing method, and a program that enable more accurate detection of abnormal cases. [Background technology]

[0002] A conventional technology is known that uses a predictive model that has been trained in advance by referencing multiple cases to predict the features contained in a case, and detects as an abnormal case those cases in which the predicted feature values ​​deviate significantly from the true values ​​of the features contained in the case.

[0003] For example, Patent Document 1 discloses a method of learning a prediction model by learning data consisting of D features as examples, using each of the features as a response variable, and using the remaining D-1 features as explanatory variables.

[0004] In Patent Document 1, D feature quantities of a case to be judged as abnormal are predicted using D trained prediction models, respectively. Then, for example, the difference between each predicted value and the true value is averaged and compared with a threshold to judge whether the case is abnormal or not.

[0005] Also, a technique has been proposed in which, when predicting one feature, a plurality of different prediction models are used to make the prediction, and the deviation degree, which indicates how far the predicted value by each prediction model deviates from the true value, is obtained (for example, see Non-Patent Document 1). According to Non-Patent Document 1, for example, the deviation degrees of the predicted values ​​by each prediction model are added together, and compared with a threshold value, to determine whether the case is abnormal or not. [Prior art documents] [Patent documents]

[0006] [Patent Document 1] U.S. Patent Application Publication No. 2005 / 0283511 [Non-patent literature]

[0007] [Non-Patent Document 1] K. Noto et. al., “FRaC: A feature-modeling approach for semi-supervised and unsupervised anomaly detection”, 2012 Summary of the Invention [Problem to be solved by the invention]

[0008] However, in a case, when a certain feature value becomes an abnormal value and many other feature values ​​also become abnormal values ​​in conjunction with the abnormal value, the feature value can be predicted from the other feature values. In such a case, the case cannot be determined as an abnormal case. This point was not taken into consideration in Patent Document 1 and Non-Patent Document 2.

[0009] One aspect of the present invention has been made in consideration of the above problems, and one example of a purpose of the present invention is to provide a technique that enables more accurate detection of abnormal cases. [Means for solving the problem]

[0010] An information processing device according to one aspect of the present invention comprises an acquisition means for acquiring cases expressed as a set of a plurality of feature quantities, a prediction means for outputting a plurality of prediction results using at least one of the feature quantities included in the cases as an objective variable and a plurality of different subsets of the plurality of feature quantities excluding the at least one of the feature quantities as explanatory variables, the subsets being subsets of the plurality of feature quantities different from each other, and an anomaly degree output means for outputting the anomaly degree of the cases by referring to the plurality of prediction results.

[0011] An information processing method according to one aspect of the present invention includes obtaining a case expressed as a set of a plurality of feature quantities, using at least one of the feature quantities included in the case as a target variable, and outputting a plurality of prediction results using a plurality of different subsets of the plurality of feature quantities excluding the at least one of the feature quantities as explanatory variables, and outputting a degree of anomaly of the case by referring to the plurality of prediction results.

[0012] A program according to one aspect of the present invention causes a computer to function as an information processing device having an acquisition means for acquiring cases expressed as a set of a plurality of feature quantities, a prediction means for outputting a plurality of prediction results using at least one of the feature quantities included in the cases as an objective variable and a plurality of different subsets of the plurality of feature quantities excluding the at least one of the feature quantities as explanatory variables, and an anomaly degree output means for outputting the anomaly degree of the cases by referring to the plurality of prediction results. Effect of the Invention

[0013] According to one aspect of the present invention, it is possible to provide a technique that enables more accurate detection of abnormal cases. [Brief description of the drawings]

[0014] [Figure 1] 1 is a block diagram showing an example of the configuration of an information processing device according to a first exemplary embodiment of the present invention. [Diagram 2] 1 is a flowchart showing a flow of an information processing method according to a first exemplary embodiment of the present invention. [Diagram 3] 1 is a block diagram showing an example of the configuration of an abnormality determination device according to an exemplary embodiment 2 of the present invention; [Figure 4] FIG. 11 is a diagram illustrating an example of case data. [Diagram 5] FIG. 13 is a diagram for explaining an outline of processing by a prediction unit. [Figure 6] 4 is a flowchart illustrating an example of an abnormality determination process performed by the abnormality determination device. [Figure 7]13 is a flowchart illustrating a detailed example of an explanatory variable specification process. [Figure 8] 13 is a flowchart illustrating a detailed example of a prediction process. [Figure 9] 13 is a flowchart illustrating a detailed example of an abnormality level output process. [Figure 10] FIG. 13 is a diagram illustrating an example of a prediction rule. [Figure 11] FIG. 11 is a block diagram showing an example of the configuration of an abnormality determination device according to an exemplary embodiment 3 of the present [Figure 12] 13 is a flowchart illustrating another example of the prediction process. [Figure 13] FIG. 13 is a diagram illustrating an example of a case where rules are extracted from a plurality of decision trees. [Figure 14] FIG. 13 is a diagram illustrating an example of a case where rules are extracted from a plurality of decision trees. [Figure 15] 13 is a flowchart illustrating yet another example of the prediction process. [Figure 16] FIG. 13 is a diagram illustrating the narrowing down of rules. [Figure 17] FIG. 11 is a block diagram showing an example of the configuration of an abnormality determination and learning device according to an exemplary embodiment 4 of [Figure 18] 11 is a flowchart illustrating an example of a learning process performed by the abnormality determination and learning device. [Figure 19] 13 is a flowchart illustrating a detailed example of a parameter update process. [Figure 20] FIG. 2 is a block diagram showing an example of a computer that executes instructions of a program, which is software that realizes each function. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0015] Example 1 A first exemplary embodiment of the present invention will be described in detail with reference to the drawings. This exemplary embodiment is a basic form of the exemplary embodiments described later.

[0016] <Overview of information processing device 20> The information processing device 20 according to this exemplary embodiment is, generally speaking, a device that predicts whether a given case, which is an example made up of a plurality of feature quantities, is abnormal or not.

[0017] More specifically, the information processing device 20, for example, An acquisition means for acquiring an example represented as a set of a plurality of feature quantities; a prediction means for outputting a plurality of prediction results using at least one of the feature quantities included in the case as a target variable, and using a plurality of different subsets of the feature quantities, which are subsets of the feature quantities excluding at least one of the feature quantities, as explanatory variables; The system further includes an abnormality degree output means for outputting the abnormality degree of the case by referring to the multiple prediction results.

[0018] <Configuration of information processing device 20> The configuration of an information processing device 20 according to this exemplary embodiment will be described with reference to Fig. 1. Fig. 1 is a block diagram showing an example of the configuration of the information processing device 20.

[0019] 1, the information processing device 20 includes an acquisition unit 21, a prediction unit 22, and an anomaly level output unit 23. The acquisition unit 21 is configured to realize an acquisition means in this exemplary embodiment. The prediction unit 22 is configured to realize a prediction means in this exemplary embodiment. The anomaly level output unit 23 is configured to realize an anomaly level output means in this exemplary embodiment.

[0020] The acquiring unit 21 acquires cases expressed as a collection of a plurality of feature amounts.

[0021] The prediction unit 22 uses at least one of the features included in the case as an objective variable, and outputs multiple prediction results using multiple different subsets of the multiple features as explanatory variables, which are subsets of the multiple features excluding at least one of the features.

[0022] For example, if an example contains D features, the subset of features is a combination of any number (including 0) of the D-1 features excluding the one feature corresponding to the objective variable.

[0023] The abnormality degree output unit 23 refers to a plurality of prediction results and outputs the abnormality degree of the case.

[0024] The anomaly degree output unit 23 obtains, for each of the multiple prediction results, a degree of deviation that indicates how far the prediction result deviates from the actual case. Then, the anomaly degree output unit 23 uses the degree of deviation of each of the multiple prediction results to calculate and output an anomaly degree that indicates the likelihood that the case is abnormal.

[0025] <Effects of the information processing device 20> According to the information processing device 20 of this exemplary embodiment, at least one of the features included in the case is used as the objective variable, and a plurality of prediction results are output using a plurality of different subsets of the features, which are subsets of the features excluding at least one of the features, as explanatory variables. Therefore, abnormal cases can be detected more accurately. In other words, since a plurality of prediction results are obtained using a plurality of different subsets as explanatory variables, even if many other features become abnormal in conjunction with a certain feature becoming an abnormal value, the degree of anomaly of the feature can be obtained.

[0026] <Flow of information processing method by information processing device 20> The flow of the information processing method executed by the information processing device 20 configured as above will be described with reference to Fig. 2. Fig. 2 is a flowchart showing the flow of the information processing method. As shown in the figure, the information processing includes steps S11 to S13.

[0027] In step S11, the acquiring unit 21 acquires a case represented as a set of a plurality of feature amounts.

[0028] In step S12, the prediction unit 22 uses at least one of the features included in the case as an objective variable, and outputs multiple prediction results using multiple different subsets of the multiple features excluding the at least one of the features as explanatory variables.

[0029] In step S13, the abnormality degree output unit 23 refers to a plurality of prediction results and outputs the abnormality degree of the case.

[0030] <Effects of information processing methods> According to the information processing method of this exemplary embodiment, at least one of the features included in the case is used as the objective variable, and a plurality of prediction results are output using a plurality of different subsets of the features, which are subsets of the features excluding at least one of the features, as explanatory variables. Therefore, abnormal cases can be detected more accurately. In other words, since a plurality of prediction results are obtained using a plurality of different subsets as explanatory variables, even if many other features also become abnormal values ​​in conjunction with a certain feature becoming an abnormal value, the degree of anomaly of the feature can be obtained.

[0031] Exemplary embodiment 2 A second exemplary embodiment of the present invention will be described in detail with reference to the drawings. Note that components having the same functions as those described in the first exemplary embodiment are given the same reference numerals, and the description thereof will be omitted as appropriate.

[0032] <Configuration of Abnormality Determination Device 10> The configuration of the abnormality determination device 10 according to this exemplary embodiment will be described with reference to Fig. 3. Fig. 3 is a block diagram showing an example of the configuration of the abnormality determination device 10. As shown in Fig. 3, the abnormality determination device 10 includes an information processing device 20, a storage unit 30, a communication unit 41, an input unit 42, and an output unit 43.

[0033] The information processing device 20 is a functional block having the same functions as the information processing device 20 described in the first exemplary embodiment.

[0034] The storage unit 30 is configured, for example, with a semiconductor memory device, etc., and stores data. In this example, the storage unit 30 stores case data and model parameters.

[0035] Here, the model parameters are parameters that define a prediction model used by the prediction unit 22 to obtain a prediction result, and may include, for example, a weighting coefficient, a threshold value, etc. referenced by the prediction model. In this exemplary embodiment, as an example, as shown in the figure, a set of multiple model parameters is stored in the storage unit 30, and these multiple model parameter sets are applied to, for example, multiple different prediction models.

[0036] The communication unit 41 is an interface for connecting the abnormality determination device 10 to a network. Although the specific configuration of the network does not limit the present exemplary embodiment, as an example, a wireless LAN (Local Area Network), a wired LAN, a WAN (Wide Area Network), a public line network, a mobile data communication network, or a combination of these networks can be used.

[0037] The input unit 42 accepts various inputs to the abnormality determination device 10. Although the specific configuration of the input unit 42 is not limited to this exemplary embodiment, as an example, the input unit 42 may be configured to include input devices such as a keyboard and a touchpad. The input unit 42 may also be configured to include a data scanner that reads data via electromagnetic waves such as infrared rays and radio waves, and a sensor that senses the environmental state.

[0038] The output unit 43 is a functional block that outputs the processing results by the abnormality determination device 10. Although the specific configuration of the output unit 43 is not limited to this exemplary embodiment, as an example, the output unit 43 is configured with a display, a speaker, a printer, etc., and displays various processing results by the abnormality determination device 10 on a screen or outputs them as sounds or figures.

[0039] The abnormality determination device 10 according to this exemplary embodiment stores a set of model parameters for a plurality of prediction models obtained by machine learning performed in advance using a plurality of cases as teacher data in the storage unit 30. The abnormality determination device 10 predicts the feature amount of a given case using the prediction model learned in advance, thereby determining whether the case is abnormal or not.

[0040] (Examples of case data and subsets) Fig. 4 is a diagram for explaining an example of case data 15. In Fig. 4, the case data 15 is data in a table format, and is made up of a plurality of records having feature quantities x1 to x4.

[0041] 4 is a case in which it is to be determined whether or not it is abnormal. As described above, the acquiring unit 21 acquires the test case 16 as a case expressed as a set of a plurality of feature amounts.

[0042] For example, in the first case in the case data 15, the value of the feature quantity x1 is −102, the value of the feature quantity x2 is −102, the value of the feature quantity x3 is −1, and the value of the feature quantity x4 is −1. In the second case in the case data 15, the value of the feature quantity x1 is −101, the value of the feature quantity x2 is −101, the value of the feature quantity x3 is −1, and the value of the feature quantity x4 is −1.

[0043] In each case of the case data 15, the value of the feature amount x1 is always equal to the value of the feature amount x2, and the value of the feature amount x3 is always equal to the value of the feature amount x4. In addition, the positive and negative signs of the values ​​of the feature amounts x1 to x4 always match. Cases that meet this condition are normal, and cases that do not meet this condition are abnormal.

[0044] For example, consider a case where a linear prediction model is trained with the value of feature quantity x1 as the objective variable and features other than feature quantity x1 as explanatory variables. In this case, the trained prediction model is x1=1*x2+0*x3+0*X4. Here, "*" is an operation symbol representing multiplication. This prediction model makes it possible to predict the value of feature quantity x1 for all cases in case data 15.

[0045] In many cases, it is difficult to obtain abnormal cases in the real world, and it is difficult to generate a predictive model that can predict that a case is abnormal from its features.

[0046] In general, if the feature values ​​predicted by a prediction model trained using a sufficiently large number of normal cases as training data are equal to or close to the feature values ​​actually contained in the case, the case is likely to be normal. This is because the correlation between the feature values ​​serving as the objective variables and the feature values ​​serving as the explanatory variables is reflected in the prediction model through learning, and the feature values ​​are predicted so as to satisfy the conditions for a normal case.

[0047] In other words, if the feature amount predicted by the prediction model deviates from the feature amount actually contained in the case, the case is highly likely to be abnormal. In this way, it is possible to determine whether or not a case is abnormal based on the deviation between the feature amount predicted by the prediction model and the feature amount actually contained in the case. In this case, even if abnormal cases are not available, a prediction model can be generated by learning normal cases.

[0048] In the test case 16, the value of the feature x1 is equal to the value of the feature x2, but the values ​​of the features x1 to x4 do not match in positive or negative sign, and therefore the test case 16 should be determined to be abnormal.

[0049] However, when the value of feature x1 is predicted using the above prediction model, x1 = 1*100+0*(-1)+0*(-1), and the value of feature x1 is correctly predicted. In this way, when feature x1 and feature x2 are abnormal at the same time for feature x3 and feature x4, a prediction using a prediction model that simply uses all feature quantities other than feature x1 as explanatory variables cannot accurately determine whether an abnormality exists.

[0050] Therefore, in this exemplary embodiment, prediction is performed using a prediction model in which a plurality of different subsets of features, which are subsets of features included in cases excluding features corresponding to objective variables, are used as explanatory variables. For example, in the case of case data 15, when the objective variable is feature x1, a prediction model that predicts feature x1 using only feature x3 is also trained.

[0051] In this case, the trained prediction model is x1=100*x3. Even if machine learning is performed on this prediction model using all of the cases included in the case data 15 as training data, the prediction error cannot be reduced to 0. In other words, the prediction errors of this prediction model will be 2, 1, 0, 1, and 2, and each prediction error will appear with approximately the same frequency. Note that a prediction model that cannot reduce the prediction error to 0 is sometimes called a weak prediction model.

[0052] When this weak prediction model is used to predict the value of feature x1 in test case 16, the result is x1 = -100, which is not a correct prediction.

[0053] In this manner, in this exemplary embodiment, a prediction model using some of the features other than the feature corresponding to the objective variable is used. In other words, by deliberately using a weak prediction model, a prediction result in which the predicted features deviate from the features actually included in the case is output, and the prediction result is used to determine whether the case is abnormal or not.

[0054] In this way, it becomes possible to appropriately determine anomalies even in cases where, for example, a certain feature value becomes abnormal and many other feature values ​​also become abnormal values ​​in conjunction with the abnormal value of the other feature values.

[0055] Here, as an example, the prediction using a linear prediction model has been described, but a prediction model of a different type from the linear prediction model may be used. For example, a prediction model such as a nonlinear SVM or a decision tree may be used. Alternatively, as described later, a prediction rule may be used instead of a prediction model.

[0056] Specifically, the feature subset and the prediction model in this exemplary embodiment are determined as follows: For example, when an example has four feature quantities, x1 to x4, and x4 is the objective variable, there are eight combinations of explanatory variables, as shown below.

[0057] [] [x1], [x2], [x3] [x1,x2], [x1, x3], [x2, x3] [x1, x2, x3] Note that [ ] indicates that 0 features included in the example are used.

[0058] According to this exemplary embodiment, the prediction unit 22 outputs a plurality of prediction results by using a prediction model corresponding to each of the subsets. In this manner, a prediction result using a weak prediction model can be obtained.

[0059] Here, each prediction model will be different depending on the combination of explanatory variables described above, but multiple prediction models using the same combination of explanatory variables may be used. For example, a prediction model of SVM with explanatory variables [x1, x2, x3] and a prediction model of decision tree with explanatory variables [x1, x2, x3] may be used. That is, one or more prediction models or prediction rules are used for each subset.

[0060] Furthermore, when determining whether an example is abnormal, prediction results of multiple feature quantities may be used. For example, if the feature quantities of an example are x1 to x8, prediction results of multiple feature quantities may be used, such as a prediction result in which x1 is the objective variable and a subset of x2 to x8 is the explanatory variable, a prediction result in which x2 is the objective variable and x1 and a subset of x3 to x8 are the explanatory variables, and so on.

[0061] In this way, it is possible to more accurately determine whether a case is abnormal or not.

[0062] Furthermore, if the correlation between the feature quantity serving as the objective variable and the feature quantity serving as the explanatory variable is taken into consideration, a more accurate judgment is possible.

[0063] For example, consider a case where multiple sensors are installed from upstream to downstream of a water pipe. In this case, if a water leak occurs upstream of the water pipe, the downstream sensor will simultaneously take an abnormal value. In this case, the feature values ​​output from the sensors are highly correlated, and it may be possible to predict each other from each feature value.

[0064] Therefore, for example, the correlation between each of the multiple features included in the case may be given in advance, and the subset used as the explanatory variables may be composed of features that have a relatively low correlation with the feature corresponding to the objective variable.

[0065] In this way, even if many features become abnormal at the same time, a prediction model can be used that can output a prediction result in which the features of the objective function deviate from the features actually included in the cases, thereby reducing the possibility that an abnormal case is mistakenly determined to be normal.

[0066] Each of the prediction models or prediction rules is defined by a set of multiple model parameters stored in the storage unit 30.

[0067] The anomaly degree output unit 23 uses the prediction result thus output to calculate and output an anomaly degree that indicates the likelihood that a given case is abnormal. Details of the calculation of the anomaly degree will be described later.

[0068] (Example of narrowing down the number of predictive models) On the other hand, running all of these machine learning predictive models would be computationally expensive, and using too many predictive models would also reduce interpretability, which is the ability to recognize ex post why a case was determined to be abnormal.

[0069] Therefore, the number of prediction models may be narrowed down. For example, the number of subsets may be limited, or the number of features constituting the subsets may be limited, thereby narrowing down the number of prediction models.

[0070] For example, the acquisition unit 21 may acquire information regarding the relevance, and the prediction unit 22 may refer to the information regarding the relevance to identify features that have a relatively low degree of relevance with at least one of the features, and form a subset using the identified features.

[0071] In this case, the features constituting the subset are limited to those with a relatively low degree of association, so the number of features is limited, and it becomes possible to narrow down the number of prediction models. For example, it is possible to select four combinations from eight combinations of features. Then, by making predictions using only prediction models in which each of the selected combinations is used as an explanatory variable, it becomes possible to narrow down the number of prediction models.

[0072] In this way, we can avoid an excessive number of predictive models, reduce computational costs, and improve interpretability.

[0073] In the above example, the number of prediction models is narrowed down by configuring explanatory variables using features that have a relatively low degree of association with the features corresponding to the objective variable, but the number of prediction models may be narrowed down using another method. For example, explanatory variables may be configured using only a predetermined number of features arbitrarily selected from features other than the features corresponding to the objective variable, and the number of prediction models may be narrowed down by repeating this selection.

[0074] That is, the prediction unit 22 may randomly select one or more feature quantities from the feature quantities excluding the feature quantity corresponding to the objective variable, and output multiple prediction results by using each of the subsets obtained by repeating the selection as an explanatory variable.

[0075] In this case, by limiting the number of times the selection is repeated, the number of subsets is limited, and the number of predictive models can be narrowed down. In this way, it is possible to prevent the number of predictive models from becoming too large, reducing computational costs and improving interpretability.

[0076] An example of an algorithm that generates a predictive model by repeatedly randomly selecting features is expressed as follows using pseudocode. Here, the algorithm is shown for a case where the number of features is L and the selection is repeated T times.

[0077] models = [] for i in range(L): models[i] = [] for j in range(T): #T is an arbitrary constant to be determined in advance features = random_select(i, L, K) #Randomly select K features from the L-1 features other than i models[i][j] = training(i, features) #Create a model that predicts feature x_i using the selected features based on the training set. In this example, the smaller the number of selections T, the fewer the number of prediction models.

[0078] (Outline of the process of the prediction unit) 5 is a diagram for explaining an outline of the process by the prediction unit 22 in the above-described exemplary embodiment. In this example, it is assumed that there are D feature amounts, Feature 1 to Feature D, as feature amounts included in the case example.

[0079] 5, the prediction unit 22 executes a prediction process using P_1 prediction models to predict feature amount 1. The prediction unit 22 also executes a prediction process using P_2 prediction models to predict feature amount 3, executes a prediction process using P_3 prediction models to predict feature amount 3, and so on, executes a prediction process using P_D prediction models to predict feature amount D.

[0080] <Flow of abnormality determination process by abnormality determination device 10> Next, an example of an abnormality determination process performed by the abnormality determination device 10 of this exemplary embodiment will be described with reference to the flowchart of FIG.

[0081] In step S31, the acquisition unit 21 acquires a case for which it is to be determined whether or not it is abnormal.

[0082] In step S32, the prediction unit 22 executes an explanatory variable specification process, thereby specifying explanatory variables for predicting a feature quantity corresponding to the objective variable in the case acquired in step S31.

[0083] Here, a detailed example of the explanatory variable specification process in step S32 in FIG. 6 will be described with reference to the flowchart in FIG.

[0084] In step S51, the prediction unit 22 collects feature amounts of the case acquired in step S31. For example, if the case includes D feature amounts, the D feature amounts are collected.

[0085] In step S52, the prediction unit 22 identifies a feature quantity that is to be a response variable. The feature quantity that is to be a response variable may be arbitrarily selected or may be specified in advance. When the feature quantity that is to be a response variable is specified in advance, it is specified by, for example, an operation accepted via the input unit 42 or predetermined information.

[0086] In step S53, the acquisition unit 21 acquires association information that is information that indicates an association between each of a plurality of feature amounts included in the case. The association information is, for example, information that describes the degree of correlation between each feature amount obtained by analyzing the case data in advance.

[0087] The acquisition unit 21 acquires the association information via, for example, the input unit 42 or the communication unit 41. Alternatively, the association information may be stored in the storage unit 30 in advance.

[0088] In step S54, the prediction unit 22 refers to the relevance information acquired in the process of step S53, and identifies feature amounts having a relatively low degree of relevance with the feature amounts of the objective variable identified in the process of step S52.

[0089] In step S55, the prediction unit 22 extracts a plurality of subsets of features using the features with relatively low relevance identified in the process of step S54. At this time, a plurality of combinations of features consisting of some or all of the features identified in the process of step S54 are extracted.

[0090] In step S56, the prediction unit 22 specifies a subset of the feature amounts extracted in the process of step S55 as explanatory variables for predicting the objective variable specified in the process of step S52. At this time, each of the multiple subsets is specified as an explanatory variable. That is, multiple explanatory variables corresponding to multiple prediction models are specified.

[0091] In this manner, the explanatory variable specification process is executed.

[0092] Here, an example has been described in which a feature quantity having a relatively low degree of correlation with the feature quantity of the objective variable is identified, and a subset that becomes an explanatory variable is extracted by the identified feature quantity. However, the subset that becomes an explanatory variable may be extracted by a different method. For example, one or more feature quantities may be randomly selected from the feature quantities excluding the feature quantity corresponding to the objective variable, and a subset obtained by repeating the selection may be extracted.

[0093] 6, after the process of step S32, the prediction unit 22 executes a prediction process in step S33, thereby predicting the dependent variable using the explanatory variables identified in the process of step S56.

[0094] Here, a detailed example of the prediction process in step S33 in FIG. 6 will be described with reference to the flowchart in FIG.

[0095] In step S71, the prediction unit 22 selects a prediction model corresponding to the subset. At this time, a prediction model corresponding to the subset of the feature amounts (i.e., explanatory variables) extracted in the process of step S55 is selected. As described above, multiple prediction models using the same combination of explanatory variables may be selected. Such multiple prediction models are defined by a set of model parameters stored in advance in the storage unit 30.

[0096] In step S72, the prediction unit 22 predicts the dependent variable using the prediction model selected in the process of step S71.

[0097] In step S73, the prediction unit 22 judges whether or not there is a next subset. If it is judged that there is a next subset, the process returns to step S71, a prediction model corresponding to the next subset is selected, and the dependent variable is predicted. In this way, a plurality of prediction results corresponding to a plurality of subsets are obtained.

[0098] If it is determined in step S73 that there is no next subset, the process proceeds to step S74, where the prediction unit 22 outputs a prediction result, thereby outputting the above-mentioned multiple prediction results.

[0099] In this manner, the prediction process is carried out.

[0100] 6, after the process of step S33, the process proceeds to step S34, where the anomaly degree output unit 23 executes an anomaly degree output process, thereby calculating and outputting the anomaly degree of the case acquired in step S31.

[0101] 3, the anomaly degree output unit 23 of this exemplary embodiment includes a probability calculation unit 101. The probability calculation unit 101 calculates a probability value p indicating the likelihood of the true value by comparing a true value, which is a value of a feature quantity corresponding to a response variable, with a prediction result. The probability value p is calculated for each of a plurality of prediction results.

[0102] Using the probability value calculated in this way, an appropriate degree of anomaly can be calculated to determine whether a case is anomalous or not.

[0103] Next, a detailed example of the abnormality degree output process in step S33 in FIG. 6 will be described with reference to the flowchart in FIG.

[0104] Here, it is assumed that the number of features x included in the case is D, and the subset of features is expressed as follows.

[0105]

number

[0106] The prediction models corresponding to each subset are represented as Ci1, Ci2, Ci3,...CiPi.

[0107] In step S91, the probability calculation unit 101 calculates a probability value p for each of the multiple prediction results output in the process of step S74.

[0108] In step S92, the abnormality degree output unit 23 calculates Surprisal, which is a parameter used to calculate the abnormality degree, based on the probability value p calculated in the process of step S91. Note that Surprisal is calculated for each of the probability values ​​p calculated in step S91. That is, Surprisal is also calculated for each of the multiple prediction results.

[0109] Surprisal is a value that represents the unexpectedness of the predicted result, and the smaller the probability value p, the larger the value becomes. Surprisal is expressed by the following formula.

[0110]

number

[0111]

number

[0112] On the other hand, the prediction results output in step S74 may be mostly small in surprise. In other words, most of the surprisals calculated in step S92 may be small. In such a case, the prediction results that could not accurately predict the objective variable will not be effectively used in the calculation of the degree of anomaly.

[0113] Therefore, for example, the degree of abnormality may be calculated using the Surprisal corresponding to the most unexpected prediction result. In this case, the degree of abnormality is calculated, for example, by the following formula.

[0114]

number

[0115] In other words, the degree of abnormality may be calculated by a calculation using k probability values ​​p (k is a natural number) of the probability values ​​p of the multiple prediction results in ascending order of value.

[0116] In this way, it is possible to calculate a more stable degree of anomaly compared to simply using the maximum value of Surprisal. In other words, since three prediction results are reflected in the degree of anomaly instead of one, it is easier to avoid, for example, the degree of anomaly changing drastically depending on the case.

[0117] Alternatively, the degree of anomaly may be calculated by weighting the number of explanatory variables used in the prediction. For example, cases in which the target variable cannot be accurately predicted in a prediction using only a small number of explanatory variables are often clearly abnormal cases at first glance.

[0118] Therefore, the degree of anomaly may be calculated by a calculation using a weighting coefficient that weights the plurality of probability values ​​so that the weight decreases as the number of explanatory variables used to obtain the prediction result corresponding to the probability value increases. In this case, the degree of anomaly may be calculated, for example, by the following formula:

[0119]

number

[0120] In this manner, the abnormality degree calculation process is executed. This also ends the abnormality degree determination process of Fig. 6. Since the abnormality degree calculated in step S93 represents the likelihood that the case is abnormal, it is possible to determine whether the case is abnormal or not by, for example, comparing the value of the abnormality degree with a threshold value.

[0121] <Effects of the abnormality determination processing device and the abnormality determination processing> In this manner, in the anomaly determination processing device and anomaly determination process according to this exemplary embodiment, prediction is performed using a prediction model in which a plurality of different subsets of feature quantities, which are subsets of feature quantities included in cases excluding feature quantities corresponding to the objective variables of the feature quantities, are used as explanatory variables. In this manner, a weak prediction model can be used intentionally, and the degree of anomaly can be calculated more appropriately. As a result, it is possible to more accurately determine whether or not a case is abnormal.

[0122] Exemplary embodiment 3 In the exemplary embodiment 2, an example has been described in which the prediction unit 22 mainly performs prediction using a prediction model. However, the prediction unit 22 may perform prediction using a prediction rule.

[0123] (Example of prediction rules extracted from a single decision tree) 10 is a diagram for explaining an example of a prediction rule in an exemplary embodiment. In the figure, an example of extracting a rule from a single decision tree that predicts a feature quantity x3 using a feature quantity x1 and a feature quantity x2 is shown.

[0124] This decision tree has a total of seven internal nodes: one at depth 0, two at depth 1, and four at depth 2. Each internal node can be considered as an IF-THEN rule.

[0125] That is, the internal node at depth 0 can be regarded as the IF-THEN rule of rule ID 0, the internal node at depth 1 can be regarded as the IF-THEN rules of rule ID 1 and rule ID 2. Furthermore, the internal node at depth 2 can be regarded as the IF-THEN rules of rule ID 3 to rule ID 6.

[0126] IF-THEN rules can be thought of as imperfect predictive models that only output a prediction if a condition is true, and no prediction otherwise.

[0127] <Configuration of Abnormality Determination Device 10> Fig. 11 is a block diagram showing a configuration example of the anomaly determination device 10 in this exemplary embodiment. As shown in the figure, in this exemplary embodiment, the prediction unit 22 has a rule extraction unit 81 that extracts rules from each node included in a decision tree trained with reference to feature quantities excluding the feature quantity corresponding to the objective variable among a plurality of feature quantities. The prediction unit 22 then outputs a plurality of prediction results using each of the extracted rules. The other configuration in Fig. 11 is the same as that in Fig. 3, and detailed description thereof will be omitted.

[0128] <Prediction process flow when using a single decision tree> Next, an example of the prediction process executed by the prediction unit 22 shown in FIG. 11 in step S33 in FIG. 6 will be described with reference to a flowchart in FIG.

[0129] In step S111, the prediction unit 22 obtains a single decision tree for predicting the dependent variable. This decision tree is, for example, a prediction model that predicts the dependent variable using all feature amounts, excluding the feature amount corresponding to the dependent variable, among the feature amounts included in the cases. The prediction model of this decision tree is defined by model parameters stored in advance in the storage unit 30.

[0130] In step S112, the rule extraction unit 81 of the prediction unit 22 extracts a rule. At this time, for example, a rule consisting of a subset of feature amounts corresponding to the explanatory variables identified by the explanatory variable identification process of step S32 described with reference to FIG. 6 is extracted.

[0131] In step S113, the prediction unit 22 predicts the dependent variable according to the rule extracted in step S112.

[0132] In step S114, the prediction unit 22 determines whether or not there is a next subset. If it is determined that there is a next subset, the process returns to step S112, where a rule corresponding to the next subset is extracted and a response variable is predicted. In this manner, a plurality of prediction results corresponding to a plurality of subsets are obtained.

[0133] If it is determined in step S112 that there is no next subset, the process proceeds to step S115, where the prediction unit 22 outputs a prediction result, thereby outputting a plurality of prediction results.

[0134] In this manner, the prediction process in this exemplary embodiment is performed. As described above, the prediction result is obtained using the rules extracted from each node included in the decision tree, so that the calculation cost is reduced.

[0135] (Example of prediction rules extracted from multiple decision trees) Here, as an example, a case has been described in which rules are extracted from a single decision tree that predicts a dependent variable using all feature quantities except for the feature quantity corresponding to the dependent variable, but rules may be extracted from multiple decision trees.

[0136] For example, as shown in Figures 13 and 14, a feature serving as an explanatory variable for predicting the feature x_i as a target variable may be randomly selected from features other than the feature x_i, so that a decision tree corresponding to each explanatory variable is used. Figures 13 and 14 are diagrams for explaining an example of extracting rules from multiple decision trees.

[0137] For example, there are feature amounts x1 to x8, and the feature amount x1 is set as the objective variable. In this case, the feature amounts that are explanatory variables are feature amounts x2 to x8, and one or more feature amounts are randomly selected from the feature amounts x2 to x8.

[0138] For example, suppose that feature amount x2, feature amount x3, and feature amount x4 are selected as combination 1, and feature amount x6, feature amount x7, and feature amount x8 are selected as combination 2. In this case, a decision tree 121A is generated corresponding to combination 1, and a decision tree 121B is generated corresponding to combination 2.

[0139] Then, IF-THEN rules are extracted from each of decision trees 121A and 121B. For example, as shown in Fig. 14, IF-THEN rule 122A is extracted from decision tree 121A, and IF-THEN rule 122B is extracted from decision tree 121B.

[0140] The IF-THEN rule 122A and the IF-THEN rule 122B extracted in this manner are combined to generate a rule set 123. The rule set 123 is a set of IF-THEN rules that use a subset of the feature quantities x2 to x8 as explanatory variables for predicting the feature quantity x1.

[0141] <Prediction process flow when using multiple decision trees> Next, an example of a prediction process in which the prediction unit 22 predicts a dependent variable using rules extracted from a plurality of decision trees will be described with reference to a flowchart in FIG.

[0142] In step S131, the prediction unit 22 selects a subset of features. At this time, for example, features to be explanatory variables are randomly selected from all features included in the cases, excluding the feature corresponding to the objective variable.

[0143] In step S132, the prediction unit 22 obtains a decision tree for predicting the dependent variable. This decision tree is a prediction model for predicting the dependent variable using the feature quantity selected in step S131. The prediction model of this decision tree is defined by model parameters stored in the storage unit 30 in advance.

[0144] In step S133, the rule extraction unit 81 of the prediction unit 22 extracts a rule from the decision tree obtained in step S132.

[0145] In step S134, the prediction unit 22 judges whether or not there is a next subset. If it is judged that there is a next subset, the process returns to step S131, a decision tree corresponding to the next subset is acquired, and a rule is extracted. In this manner, a rule is extracted from each of the multiple decision trees.

[0146] If it is determined in step S134 that there is no next subset, the process proceeds to step S135.

[0147] In step S135, the prediction unit 22 generates a rule set by combining the rules extracted in step S133.

[0148] In step S136, the prediction unit 22 predicts the feature quantity that is the objective variable by using the rule set generated in the process of step S135. As a result, a plurality of prediction results are output corresponding to each of the rules included in the rule set.

[0149] In this manner, the prediction process is carried out.

[0150] (Example of narrowing down prediction rules) 13 and 14, prediction is performed using a rule set obtained by combining rules from each of a plurality of decision trees as is. However, the rule set may contain redundant rules or rules that are not very useful for prediction.

[0151] The prediction unit 22 may, for example, exclude redundant rules and rules that are not very useful for prediction, and narrow down the rules included in the rule set 123 to generate a narrowed down rule set 124, as shown in FIG.

[0152] As an example of a method for narrowing down rules, you can use the technique described in "Yuzuru Okajima, Kunihiko Sadamasa. Decision List Optimization based on Continuous Relaxation: The SIAM International Conference on Data Mining (SDM19), 2019." This technique uses stochastic gradient decent (SGD) to input a set of rules and output a list of rules (decision list) consisting of only those rules that are useful for prediction.

[0153] In this way, the prediction unit 22 may select a number of rules to be used for outputting a prediction result by narrowing down the extracted rules to a small number of rules by excluding redundant rules. Narrowing down the number of rules further reduces the calculation cost.

[0154] An example of an algorithm for generating a set of narrowing down rules is expressed as follows using pseudocode: Here, it is assumed that there are D features included in the case example, and a set of prediction rules R_i is obtained for feature x_i. rules = [] for i in [1, ..., D]: ruleset_i = [] # A set of rules R_i that predicts x_i for j in [1, ..., T]: #T is an arbitrary constant determined in advance features = random_select(i, D, K) #Randomly select K features from D-1 features other than i tree = training(i, features) #Create a decision tree based on the training set that predicts feature x_i using the selected features. rules = get_rules_from_tree(tree) # Extract rules from a decision tree ruleset_i.append(rules) #Add the extracted rules to R_i rulesets[i] = filter(ruleset_i) # Filter rule sets An example of an algorithm for calculating the degree of anomaly for a given example x using a set of narrowing down rules is expressed in pseudocode as follows: Here, a vector with D elements, x=[x1, x2, ...], is input as an example. surprisals = [] for i in [1, ..., D]: #For D features rules = get_applicable(rulesets[i]) # Extract rules that x_i applies to from rule set R_i error = get_error(rules.predict, x_i) #Calculate the error between the rule prediction and x_i surprisals[i] = get_surprisal(error) #Calculate the surprisal of the feature from the error score = get_anomaly_score(surprisals) #Calculate the anomaly score of an example from the surprisals of D features Exemplary embodiment 4 Next, a fourth exemplary embodiment of the present invention will be described in detail with reference to the drawings. Note that components having the same functions as those described in the first to third exemplary embodiments are given the same reference numerals, and the description thereof will be omitted as appropriate.

[0155] <Configuration of Abnormality Judgment and Learning Device 10A> The configuration of an abnormality determination and learning device 10A according to this exemplary embodiment will be described with reference to Fig. 17. The abnormality determination and learning device 10A is a device further having a function of learning a set of model parameters in the storage unit 30 in the abnormality determination device 10.

[0156] As shown in the figure, the abnormality determination and learning device 10A has an information processing device 20A. Similar to the information processing device 20 shown in Fig. 3, the information processing device 20A has an acquisition unit 21, a prediction unit 22, and an abnormality degree output unit 23, and further has a teacher data acquisition unit 24 and a learning unit 25. The other configuration of the abnormality determination and learning device 10A is similar to that of the abnormality determination device 10 shown in Fig. 3.

[0157] The teacher data acquisition unit 24 included in the information processing device 20A acquires teacher data by referring to the case data. The learning unit 25 performs machine learning using the teacher data acquired by the teacher data acquisition unit 24 to learn a prediction model.

[0158] <Flow of learning process by abnormality determination / learning device 10A> FIG. 18 is a flowchart illustrating an example of the learning process performed by the abnormality determination and learning device 10A.

[0159] In step S151, the teacher data acquisition unit 24 acquires examples. At this time, a plurality of examples to be the subject of learning are acquired.

[0160] In step S152, the teacher data acquisition unit 24 executes an explanatory variable specification process. This specifies explanatory variables used to predict the objective variable. This process is similar to the process described with reference to Fig. 7, so a detailed description will be omitted. The teacher data acquisition unit 24 may control the prediction unit 22 to execute the explanatory variable specification process of step S152.

[0161] In the explanatory variable specification process in step S152, as described above with reference to FIG. 7, a plurality of explanatory variables corresponding to a plurality of prediction models are specified.

[0162] In addition, in step S152, explanatory variables are identified for each of the multiple cases acquired in step S151.

[0163] In step S153, the teacher data acquiring unit 24 and the learning unit 25 execute a parameter update process.

[0164] Here, a detailed example of the parameter update process in step S153 in FIG. 18 will be described with reference to the flowchart in FIG.

[0165] In step S171, the teacher data acquisition unit 24 acquires teacher data consisting of a set of features corresponding to the objective variables and features corresponding to the explanatory variables. At this time, for example, the features corresponding to the objective variables identified in step S52 of Fig. 7 and the features corresponding to the explanatory variables identified in step S56 are extracted from the examples acquired in step S151 of Fig. 18 to be used as teacher data.

[0166] It should be noted that multiple pieces of teacher data are acquired corresponding to the multiple cases acquired in step S151.

[0167] In step S172, the learning unit 25 calculates each component of the weight matrix using the teacher data acquired in step S171. At this time, for example, in order to predict the objective variable, each component of the weight matrix to be multiplied with the explanatory variable is calculated. At this time, each component of the weight matrix to be multiplied with each of a plurality of explanatory variables is calculated. Each component of the calculated weight matrix becomes a set of model parameters.

[0168] In this example, the objective variable is calculated using a linear prediction model, but other types of prediction models may be used. In short, it is sufficient that a set of model parameters of the prediction model is calculated in step S172.

[0169] In step S173, the learning unit 25 updates the model parameters.

[0170] The processes of steps S172 and S173 are executed multiple times according to the number of teacher data acquired in step S171, and the model parameters are repeatedly updated. In this manner, the model parameters are updated, and the learning of the prediction model progresses.

[0171] In this manner, the parameter update process is executed. When the parameter update process is completed, the learning process of FIG. 18 is completed, and a set of model parameters of a plurality of prediction models used to predict one dependent variable is obtained. The learning process is executed a plurality of times by changing the dependent variable. As a result, a set of model parameters of the prediction models used in the prediction process executed in step S33 of FIG. 6 is obtained.

[0172] <Effects of the abnormality judgment / learning device and learning process> In this way, the anomaly determination and learning device and learning process according to this exemplary embodiment can obtain model parameters of a prediction model used in a prediction process. In addition, the prediction process can be executed using a prediction model of the obtained model parameters, and it can be determined whether or not a case is abnormal.

[0173] [Software implementation example] Some or all of the functions of the information processing device 20, the abnormality determination device 10, and the abnormality determination and learning device 10A may be realized by hardware such as an integrated circuit (IC chip), or may be realized by software.

[0174] In the latter case, the information processing device 20, the abnormality determination device 10, and the abnormality determination and learning device 10A are realized by, for example, a computer that executes instructions of a program, which is software that realizes each function. An example of such a computer (hereinafter, referred to as computer C) is shown in FIG.

[0175] The computer C includes at least one processor C1 and at least one memory C2. The memory C2 stores a program P for operating the computer C as the information processing device 20, the abnormality determination device 10, or the abnormality determination and learning device 10A. In the computer C, the processor C1 reads and executes the program P from the memory C2 to realize each function of the information processing device 20, the abnormality determination device 10, or the abnormality determination and learning device 10A.

[0176] The processor C1 may be, for example, a central processing unit (CPU), a graphic processing unit (GPU), a digital signal processor (DSP), a micro processing unit (MPU), a floating point number processing unit (FPU), a physics processing unit (PPU), a microcontroller, or a combination of these. The memory C2 may be, for example, a flash memory, a hard disk drive (HDD), a solid state drive (SSD), or a combination of these.

[0177] The computer C may further include a RAM (Random Access Memory) for expanding the program P during execution and for temporarily storing various data. The computer C may further include a communication interface for transmitting and receiving data to and from other devices. The computer C may further include an input / output interface for connecting input / output devices such as a keyboard, a mouse, a display, and a printer.

[0178] Furthermore, the program P can be recorded on a non-transitory tangible recording medium M that can be read by the computer C. Such a recording medium M can be, for example, a tape, a disk, a card, a semiconductor memory, or a programmable logic circuit. The computer C can obtain the program P via such a recording medium M. Furthermore, the program P can be transmitted via a transmission medium. Such a transmission medium can be, for example, a communication network or broadcast waves. The computer C can also obtain the program P via such a transmission medium.

[0179] [Additional Note 1] The present invention is not limited to the above-described embodiment, and various modifications are possible within the scope of the claims. For example, embodiments obtained by appropriately combining the technical means disclosed in the above-described embodiment are also included in the technical scope of the present invention.

[0180] [Additional Note 2] Some or all of the above-described embodiments can be described as follows. However, the present invention is not limited to the following described aspects.

[0181] (Appendix 1) An acquisition means for acquiring an example represented as a set of a plurality of feature quantities; a prediction means for outputting a plurality of prediction results using at least one of the feature quantities included in the case as a response variable and using a plurality of different subsets of the feature quantities, which are subsets of the feature quantities excluding the at least one of the feature quantities, as explanatory variables; an anomaly level output means for outputting an anomaly level of a case by referring to the plurality of prediction results; 23. An information processing apparatus comprising:

[0182] (Appendix 2) The prediction means includes: Outputting a plurality of the prediction results by using one or more prediction models or prediction rules for each of the subsets. 2. The information processing device according to claim 1 .

[0183] (Appendix 3) Correlations between the plurality of feature quantities are given in advance, The subset is composed of feature quantities having a relatively low degree of association with at least one of the feature quantities. 3. The information processing device according to claim 1 or 2.

[0184] (Appendix 4) The acquiring means further acquires information regarding the relevance, The prediction means includes: By referring to the information on the relevance, a feature quantity having a relatively low relevance with at least one of the feature quantities is identified, and the subset is formed by the identified feature quantities. 4. The information processing device according to claim 3.

[0185] (Appendix 5) The prediction means includes: randomly selecting one or more feature quantities from the feature quantities excluding at least one of the feature quantities; Each of the subsets obtained by repeating the selection is used as the explanatory variables to output a plurality of the prediction results. 3. The information processing device according to claim 2.

[0186] (Appendix 6) The prediction means includes: a rule extraction means for extracting a rule from each node included in a trained decision tree by referring to feature quantities excluding at least one of the feature quantities among the plurality of feature quantities, A plurality of the prediction results are output using each of the extracted rules. 3. The information processing device according to claim 2.

[0187] (Appendix 7) The rule extraction means A plurality of rules to be used for outputting the prediction result is selected by narrowing down the rules extracted from each node of the decision tree to a small number of rules by excluding redundant rules. 7. The information processing device according to claim 6,

[0188] (Appendix 8) the anomaly level output means includes a probability calculation means for calculating a probability value representing a likelihood of a true value by comparing a true value, which is a value of a feature included in the case and corresponds to the objective variable, with the prediction result; The degree of abnormality is calculated by a calculation using a plurality of the probability values ​​corresponding to the plurality of prediction results. 8. An information processing device according to any one of claims 1 to 7.

[0189] (Appendix 9) The abnormality level is The information processing device according to claim 8, wherein the probability value is calculated by an operation using m probability values ​​(m is a natural number) in ascending order of value among the plurality of probability values.

[0190] (Appendix 10) The abnormality level is The plurality of probability values ​​are calculated by a weighting coefficient that weights the plurality of probability values ​​so that the weight decreases as the number of explanatory variables used to obtain the prediction result corresponding to the plurality of probability values ​​increases, and the plurality of probability values ​​are calculated by a calculation using the weighting coefficient and the plurality of probability values. 9. The information processing device according to claim 8,

[0191] (Appendix 11) Obtaining examples represented as a set of multiple features; outputting a plurality of prediction results using at least one of the feature quantities included in the case examples as a target variable, and using a plurality of different subsets of the plurality of feature quantities as explanatory variables, the subsets being a subset of the plurality of feature quantities excluding the at least one of the feature quantities; and outputting the degree of anomaly of the case by referring to the plurality of prediction results. 23. An information processing method comprising:

[0192] (Appendix 12) Computer, An acquisition means for acquiring an example represented as a set of a plurality of feature quantities; a prediction means for outputting a plurality of prediction results using at least one of the feature quantities included in the case as a response variable and using a plurality of different subsets of the feature quantities, which are subsets of the feature quantities excluding the at least one of the feature quantities, as explanatory variables; and an abnormality level output means for outputting an abnormality level of a case by referring to the plurality of prediction results. A program characterized by:

[0193] [Additional Note 3] A part or all of the above-described embodiments can be further expressed as follows.

[0194] At least one processor, the processor comprising: An acquisition process for acquiring examples represented as a set of a plurality of feature quantities; a prediction process for outputting a plurality of prediction results using at least one of the feature quantities included in the examples as a response variable and using a plurality of different subsets of the feature quantities, which are subsets of the feature quantities excluding the at least one of the feature quantities, as explanatory variables; an anomaly level output process for outputting an anomaly level of an instance by referring to the plurality of prediction results; Execute.

[0195] The information processing device may further include a memory that stores a program for causing the processor to execute the acquisition process and the output sequence generation process. The program may be recorded on a computer-readable, non-transitory, tangible recording medium. [Explanation of symbols]

[0196] 10 Abnormality determination device 10A Abnormality Judgment and Learning Device 20 Information processing device 20A Information processing equipment 21 Acquisition Department 22 Prediction Department 23 Abnormality output section 24 Teacher Data Acquisition Department 25 Learning Department 30 Storage section 41 Communications Department 42 Input section 43 Output section 81 Rule Extraction Unit 101 Probability Calculation Department

Claims

1. An acquisition means for acquiring an example represented as a set of a plurality of feature quantities; a prediction means for outputting a plurality of prediction results using at least one of the feature quantities included in the case as a response variable and using a plurality of different subsets of the feature quantities, which are subsets of the feature quantities excluding the at least one of the feature quantities, as explanatory variables; an anomaly level output means for outputting an anomaly level of a case by referring to the plurality of prediction results; 23. An information processing apparatus comprising:

2. The prediction means includes: Outputting a plurality of the prediction results by using one or more prediction models or prediction rules for each of the subsets.

2. The information processing apparatus according to claim 1,

3. Correlations between the plurality of feature quantities are given in advance, The subset is composed of feature quantities having a relatively low degree of association with at least one of the feature quantities.

3. The information processing apparatus according to claim 1, wherein the information processing apparatus is a computer.

4. The acquiring means further acquires information regarding the relevance, The prediction means includes: By referring to the information on the relevance, a feature quantity having a relatively low relevance with at least one of the feature quantities is identified, and the subset is formed by the identified feature quantities.

4. The information processing apparatus according to claim 3.

5. The prediction means includes: randomly selecting one or more feature quantities from the feature quantities excluding at least one of the feature quantities; Each of the subsets obtained by repeating the selection is used as the explanatory variables to output a plurality of the prediction results.

3. The information processing apparatus according to claim 2.

6. The prediction means includes: a rule extraction means for extracting a rule from each node included in a trained decision tree by referring to feature quantities excluding at least one of the feature quantities among the plurality of feature quantities, A plurality of the prediction results are output using each of the extracted rules.

3. The information processing apparatus according to claim 2.

7. The rule extraction means A plurality of rules to be used for outputting the prediction result is selected by narrowing down the rules extracted from each node of the decision tree to a small number of rules by excluding redundant rules.

7. The information processing apparatus according to claim 6,

8. the anomaly level output means includes a probability calculation means for calculating a probability value indicating a likelihood of a true value by comparing a true value, which is a value of a feature included in the case and corresponds to the objective variable, with the prediction result; The degree of abnormality is calculated by a calculation using a plurality of the probability values ​​corresponding to the plurality of prediction results. The information processing device according to claim 1 .

9. Obtaining examples represented as a set of multiple features; outputting a plurality of prediction results using at least one of the feature quantities included in the case as a target variable, and using a plurality of different subsets of the plurality of feature quantities as explanatory variables, the subsets being a subset of the plurality of feature quantities excluding the at least one of the feature quantities; and outputting the degree of anomaly of the case by referring to the plurality of prediction results.

23. An information processing method comprising:

10. Computer, An acquisition means for acquiring an example represented as a set of a plurality of feature quantities; a prediction means for outputting a plurality of prediction results using at least one of the feature quantities included in the case as a response variable and using a plurality of different subsets of the feature quantities, which are subsets of the feature quantities excluding the at least one of the feature quantities, as explanatory variables; and an abnormality level output means for outputting an abnormality level of a case by referring to the plurality of prediction results. A program characterized by:

Citation Information

Patent Citations

  • Determination device, determination method and determination program

    JP2018027577A

  • Information processing device, information processing method, and program

    JP2019061494A

  • Cross-feature analysis

    US20050283511A1