Prediction method, prediction device, and computer readable recording medium

By spanning tree structure data and determining the contribution degree of attribute values, the problem of increasing processing costs caused by the rotational attempt of multiple classification rules in the prior art is solved, and an efficient prediction process is achieved.

CN112183571BActive Publication Date: 2024-11-19FUJITSU LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202010535196.X
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Priority Date
2019-07-05
Filing Date
2020-06-12
Publication Date
2024-11-19
Estimated Expiration
2040-06-12

AI Technical Summary

Technical Problem

The prior art attempts in turn based on multiple classification rules when predicting the best action, resulting in increased processing costs.

Method used

By spanning tree structure data, multiple rule information is aggregated based on predetermined order conditions of multiple attributes, and the contribution degree of attribute values ​​is determined, thereby determining the most likely predetermined tag value in the input data, thereby predicting the best action.

Benefits of technology

It effectively reduces processing costs, improves prediction efficiency, avoids the process of taking turns to try various classification rules, and improves the accuracy and efficiency of prediction.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN112183571B_ABST
    Figure CN112183571B_ABST
Patent Text Reader

Abstract

A computer-readable recording medium storing a prediction program, a prediction method, and a prediction device are provided. The prediction program according to the embodiment enables a computer to perform receiving processing, generating processing, and determining processing. The receiving processing receives input data to be predicted. Tree structure data is generated based on multiple rule information and according to predetermined sequence conditions for multiple attributes, each rule information is indicated by the association between a combination of attribute values ​​of the multiple attributes and a label, and the tree structure data is obtained by aggregating the multiple rule information, and the tree structure data includes attribute values ​​as branches. The determination processing determines the contribution of making the determination result of the predetermined value of the predetermined attribute reach the predetermined label based on the possibility of obtaining the value of the predetermined label as the determination result of the predetermined value of the predetermined attribute when the attribute value of the predetermined attribute among the multiple attributes in the tree structure data is determined as a predetermined value based on the input data.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The embodiments discussed herein relate to a prediction program, a prediction method, and a prediction device. Background Art

[0002] As a technique for nonlinear classification of discrete data, there is a technique of performing machine learning on a classification rule (ie, a decision tree) for classifying data using supervised training data and predicting a classification result of input data using the learned decision tree.

[0003] Non-patent document 1: L. Breiman, Machine Learning, Vol. 45, pp. 5-32 (2001).

[0004] One of the purposes of predicting input data is to specify (predict) the best action using classification rules (eg, how to control the next process in a manufacturing process, what action to perform next for marketing to a customer, etc.).

[0005] However, a plurality of classification rules for prediction may be generated. For this reason, the above-described related art has a problem in that, when predicting the best action, all actions based on each of the plurality of classification rules are tried (in turn), and processing costs are increased.

[0006] Therefore, an object of one aspect of an embodiment of the present invention is to provide a prediction program, a prediction method, and a prediction device that can efficiently perform prediction of input data. Summary of the invention

[0007] According to one aspect of the embodiment, a computer-readable recording medium stores a prediction program that enables a computer to perform processing, the processing comprising: receiving input data to be predicted; generating tree structure data based on multiple rule information and according to predetermined sequence conditions for multiple attributes, each rule information in the multiple rule information being indicated by an association between a combination of attribute values ​​of the multiple attributes and a label, the tree structure data being obtained by aggregating the multiple rule information, the tree structure data including attribute values ​​as branches; and when an attribute value of a predetermined attribute among multiple attributes in the tree structure data is determined as a predetermined value based on the input data, determining a contribution of causing the determination result of the predetermined value of the predetermined attribute to reach a predetermined label based on the possibility of obtaining a value of a predetermined label as a determination result of the predetermined value of the predetermined attribute. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] Figure 1 is a block diagram showing a functional configuration example of an information processing device according to an embodiment;

[0009] Figure 2is a flowchart showing an operation example of an information processing apparatus according to an embodiment;

[0010] Figure 3 is an explanatory diagram showing an example of training data;

[0011] Figure 4 It is an illustration diagram used to illustrate the generation of a hypothesis.

[0012] Figure 5 is an explanatory diagram used to illustrate the generation of hypotheses;

[0013] Figure 6 is an explanatory diagram used to illustrate the generation of hypotheses;

[0014] Figure 7 is an explanatory diagram showing an example of generated hypotheses;

[0015] Figure 8 is an explanatory diagram used to illustrate the hypothesis matching with the input data;

[0016] Fig. 9 is an illustrative diagram for explaining weighting by logistic regression;

[0017] Fig.10 is an illustrative diagram used to illustrate the hypothetical choice;

[0018] Fig.11 is an explanatory diagram for explaining the generation of tree structure data;

[0019] Fig. 12A is an explanatory diagram for explaining a calculation example of the probability of obtaining a value of a predetermined label;

[0020] Fig. 12B is an explanatory diagram for explaining a calculation example of the probability of obtaining a value of a predetermined label;

[0021] Fig. 12C is an explanatory diagram for explaining a calculation example of the probability of obtaining a value of a predetermined label;

[0022] Fig.13 is an explanatory diagram illustrating contribution degrees;

[0023] Fig.14A is an explanatory diagram for explaining a calculation example of the probability of obtaining a value of a predetermined label;

[0024] Fig. 14B is an explanatory diagram for explaining a calculation example of the probability of obtaining a value of a predetermined label;

[0025] Fig. 14C is an explanatory diagram for explaining a calculation example of the probability of obtaining a value of a predetermined label;

[0026] Fig.15 is an explanatory diagram for explaining the output of the prediction result; and

[0027] Fig.16 is an explanatory diagram for explaining a hardware configuration example of an information processing device according to an embodiment. DETAILED DESCRIPTION

[0028] The preferred embodiments will be described with reference to the accompanying drawings. In the embodiments, configurations having the same functions are represented by the same reference numerals, and redundant descriptions thereof will be omitted. Note that the prediction program, prediction method, and prediction device described in the following embodiments are merely examples, and the embodiments are not limited thereto. In addition, each of the following embodiments may be appropriately combined within a non-contradictory range.

[0029] Figure 1 is a block diagram showing a functional configuration example of an information processing device according to an embodiment.

[0030] like Figure 1 As shown, the information processing device 1 includes an input unit 10, a storage unit 20, a hypothesis generating unit 30, a learning unit 40, a prediction unit 50, and an output unit 60.

[0031] The input unit 10 is a processing unit that receives input of various data such as training data 21 related to machine learning and input data 22 to be predicted. The input unit 10 stores the training data 21 or input data 22 received as input in the storage unit 20.

[0032] The storage unit 20 stores various data, such as training data 21 , input data 22 , hypothesis set data 23 , weight data 24 , and result data 25 .

[0033] The hypothesis generating unit 30 comprehensively searches for each hypothesis including a combination of the explanatory variables (a rule (basis) explaining that prediction is made based on the target variables) from the training data 21 having the explanatory variables and the target variables.

[0034] Next, the hypothesis generation unit 30 classifies any training data 21 based on the explanatory variables and target variables of the training data 21 for each of the hypotheses searched, and specifies the hypothesis that meets the specific condition. Here, the specific condition is that the number or proportion of the training data 21 classified into a predetermined category by the rule indicated by the hypothesis (combination of explanatory variables) is equal to or greater than a predetermined value. For example, the hypothesis generation unit 30 specifies the searched hypothesis based on a certain number or more of samples and (or) a certain proportion or greater of samples, and the searched hypothesis indicates that the number or proportion of the training data 21 classified by the hypothesis is equal to or greater than a predetermined value and the classification result of the hypothesis belongs to a specific category. That is, the hypothesis generation unit 30 specifies the hypothesis that can correctly explain that the prediction is based on the target variable of the training data 21.

[0035] Next, the hypothesis generation unit 30 adds the specified hypothesis to the hypothesis set. In this way, the hypothesis generation unit 30 lists the hypothesis that can correctly explain that the prediction is made based on the target variable of the training data 21 in the hypothesis set. Next, the hypothesis generation unit 30 stores the hypothesis set data 23 indicating the hypothesis set in which the hypothesis is listed in the storage unit 20. Each hypothesis listed in the hypothesis set is an example of a plurality of rule information, each of which is indicated by the association of a combination of attribute values ​​of a plurality of attributes (combination of explanation variables) with a label (target variable).

[0036] The learning unit 40 performs learning to calculate the weight of each of the plurality of hypotheses based on whether each of the plurality of hypotheses included in the hypothesis set of the hypothesis set data 23 is established for each training data 21. The learning unit 40 stores the weight of each of the plurality of hypotheses obtained by the learning result as the weight data 24 in the storage unit 20. The hypothesis set data 23 and the weight data 24 obtained in this way are a prediction model for obtaining a prediction result.

[0037] The prediction unit 50 is a processing unit that generates a prediction result based on the input data 22 to be predicted using a hypothesis set based on the hypothesis set data 23 and weights (i.e., a prediction model) of a plurality of hypotheses based on the weight data 24. The prediction unit 50 stores the generated prediction result as result data 25 in the storage unit 20.

[0038] For example, the input data 22 includes known actions (part of the explanatory variables) and target labels (target variables). For unknown actions (remaining explanatory variables), the prediction unit 50 uses a hypothesis set based on the hypothesis set data 23 and a prediction model to predict the best explanatory variable value as the target label after performing the known action, that is, the best action.

[0039] For example, in a manufacturing process, in the case of predicting how to control the next process so as to manufacture a non-defective product, the known actions included in the input data 22 include observation values, control setting values, etc. in the manufacturing process. In addition, the target label includes a label indicating that the product manufactured in the manufacturing process is a non-defective product. Thus, the prediction unit 50 can predict how to control the next process (unknown action) to manufacture a non-defective product.

[0040] In addition, for example, in the case of predicting what action needs to be performed for a customer in order to successfully conduct marketing, the known action included in the input data 22 includes the content received from the user in the marketing. In addition, the target tag includes a tag indicating that the marketing was successful. Thus, the prediction unit 50 can predict what action (unknown action) needs to be performed for a customer in order to successfully conduct marketing.

[0041] Specifically, the prediction unit 50 obtains data by aggregating each hypothesis listed in the hypothesis set according to a predetermined order condition for a plurality of attributes based on the hypothesis set data 23 , and generates tree structure data including attribute values ​​(explanatory variables) as branches.

[0042] It is assumed that the order conditions of the plurality of attributes are preset as user settings in the rule base or the input data 22. Specifically, an order corresponding to the processing order in the manufacturing process or the like (e.g., A→B→C→...) is preset for the plurality of attributes. Therefore, the prediction unit 50 generates tree structure data having attribute values ​​(e.g., observed values ​​or control setting values ​​in the manufacturing process) as branches in an order corresponding to the process processing in the manufacturing process or the like.

[0043] Next, the prediction unit 50 uses the tree structure indicated by the generated tree structure data to search for the best explanatory variable value, that is, the best action, which is the target label after executing the known action included in the input data 22. Specifically, when the attribute value of a predetermined attribute among the multiple attributes in the tree structure data is determined as a predetermined value based on the input data 22, the prediction unit 50 determines the contribution of making the predetermined value of the predetermined attribute reach the predetermined label based on the possibility (for example, the prediction score) of obtaining the value of the predetermined label as the determination result of the prediction model. Next, the prediction unit 50 searches for the attribute value in the predetermined attribute based on the determined contribution. Specifically, the value that makes the contribution higher is set as the attribute value in the predetermined attribute.

[0044] For example, the prediction unit 50 performs pruning by applying known attribute values ​​included in the input data 22 to the tree structure data. Next, the prediction unit 50 determines a contribution for the fact that, for an attribute whose attribute value is unknown among a plurality of attributes, a determination result when the unknown attribute value of the attribute is determined to be a predetermined value is a label included in the input data 22. Next, the prediction unit 50 searches for attribute values ​​in the unknown attributes based on the determined contribution. For example, the following method is found: by enabling the prediction unit 50 to compare the contribution determined according to a branch having an attribute value of 0 based on the tree structure data with the contribution determined according to a branch having an attribute value of 1 based on the tree structure data, the contribution is made higher.

[0045] The output unit 60 is a processing unit that reads out the result data 25 stored in the storage unit 20 and outputs the read result data 25 to a display, a file, etc. Thus, the information processing device 1 outputs the prediction result predicted by the prediction unit 50 to a display, a file, etc. Specifically, the information processing device 1 outputs the prediction result of the optimal action predicted by the prediction unit 50.

[0046] As described above, the information processing device 1 is an example of a learning device and a prediction device. In the present embodiment, a configuration is shown in which one information processing device 1 performs learning and prediction in an integrated manner, but an independent information processing device 1 can also realize learning and prediction.

[0047] Next, the processing of each of the above functional units will be described in detail while explaining an operation example of the information processing device 1 . Figure 2 is a flowchart showing an operation example of the information processing device 1 according to the embodiment.

[0048] like Figure 2 As shown, the operation of the information processing device 1 includes an operation (S1) for generating a prediction model during learning and a prediction operation (S2) for obtaining a prediction result by applying the generated prediction model to the input data 22 to be predicted. First, the operation (S1) during learning will be described.

[0049] like Figure 2 As shown, when the process starts, the input unit 10 receives an input of training data 21 ( S11 ) and stores the input in the storage unit 20 .

[0050] Figure 3 21 is an explanatory diagram showing an example of training data. The training data 21 is supervisory data for each of a plurality of cases, and includes explanatory variables A to D that explain the nature of the data and a target variable of a classification result (correct answer information) classified as + or -.

[0051] like Figure 3 As shown, the training data (P 1 To P 4 and N 1 To N 3 ) includes explanatory variables A to D (information for prediction) that explain the nature of the data and a target variable (information to be predicted) that is information about the correct answer indicating the category (classification) as + or -. For example, the training data P 1 To P 4 is the data where each of the target variables A to D is 0 or 1 and is classified as +. Similarly, the training data N 1 To N 3 It is data in which each of the target variables A to D is 0 or 1 and is classified as -.

[0052] For example, in the manufacturing process domain, in the training data (P 1 To P 4 and N 1 To N 3 ) is used to generate a prediction model for classifying the results of manufactured products (non-defective products / defective products) based on processing data, the explanatory variables A to D correspond to the observed values, control values, etc. of each processing. In addition, the target variable corresponds to the manufacturing result, for example, non-defective product (P) / defective product (N).

[0053] ˉ

[0054] The variable (1 / 0) is indicated by the presence or absence of an overline (hereinafter, represented as "bar"). For example, A indicates A=1, and Indicates A = 0. In addition, the target variable (+ / -) is represented by a shaded pattern. For example, the training data P 1 To P 4 The shaded pattern indicates that the target variable is +. In addition, the training data N 1 To N 3 The shading pattern of indicates that the target variable is -. Note that these representations are common to the other figures.

[0055] Next, it is assumed that the generation unit 30 is for the training data (P 1 To P 4 and N 1 To N 3 ) each of the explanatory variables included in the comprehensive list of possible value combinations (unused = *, value = 1, value = 0), that is, assumptions (S12).

[0056] Note that the number of explanation variables to be combined can be limited (constrained) to be equal to or less than a predetermined number. For example, in the case of four explanation variables A to D, the number of explanation variables to be combined can be limited to two or less (combining at least two of the four explanation variables of "unused = *"). Therefore, an increase in the number of combinations can be suppressed in advance.

[0057] Next, it is assumed that the generation unit 30 selects a predetermined combination from the combinations listed in S12 (S13). Next, it is assumed that the generation unit 30 selects a predetermined combination based on the training data (P 1 To P 4 and N 1 To N 3 ) of the explanatory variables and the target variable to classify the selected combination into training data (P 1 To P 4 and N 1 To N 3 ) and determines whether the selected combination is a valid combination that satisfies a specific condition (S14).

[0058] Figure 4 It is an explanatory diagram showing the generation of a hypothesis. Figure 4 As examples, a combination C01 in which all of the four explanatory variables A to D are "*" to a combination C09 which is CD (AB is "*") are shown.

[0059] like Figure 4 As shown, it is assumed that the generation unit 30 is based on the training variables (P 1 To P 4 and N 1 To N 3 ) lists the training data (P) corresponding to the hypothesis (rule) of each combination C01 to C09. 1 To P 4 and N 1 To N 3 ).

[0060] For example, the combination C02 The rule (the remaining three explanatory variables are "unused = *") corresponds to the training data P 2 、N 1 and N 2 In this combination C02 rules The target variable is the training data of + (P 2 ) and its target variable is - training data (N 1 and N 2 ) mixture. Therefore, combination C02 cannot be a hypothesis that correctly describes that the combination is classified into a specific category, and is not a valid combination.

[0061] Here, the target variable is the training data (P 1 , P 3 and P 4 ) corresponds to the rule for combination C04 That is, combination C04 is likely to correctly describe the training data classified as + category (P 1 , P 3 and P 4 ) is equal to or greater than a predetermined value and is classified as a + category rule. Therefore, it is assumed that the generation unit 30 determines that the combination C04 is a valid combination (hypothesis) classified as the + category. Similarly, the hypothesis generating unit 30 also determines that the combinations C05 and C06 are valid combinations (hypothesis) classified as the + category.

[0062] In addition, the rules for combining C08 Corresponding to the training data whose target variable is -(N 1 and N 2 That is, combination C08 is likely to correctly describe the training data (N 1 and N 2 ) is equal to or greater than a predetermined value and is classified as a rule of - category. Therefore, it is assumed that the generation unit 30 determines that the combination C08 is a valid combination classified as -category (hypothesis).

[0063] The training data (P 1 To P 4 and N 1 To N 3 ), which is a condition for determining a valid combination. For example, since noise may be mixed in the training data, the training data may be set to allow a predetermined number of categories (e.g., -) that are opposite to a predetermined category (e.g., +).

[0064] As an example, when the noise corresponding to one training data is allowed, combination C03 (D) is determined to be a valid combination (hypothesis) classified as the + category. Similarly, combination C07 (C) is determined to be a valid combination (hypothesis) classified as the - category.

[0065] Refer to Figure 2 , if the combination is invalid (S14: No), the hypothesis generation unit 30 proceeds to S17 without adding the selected combination to the hypothesis set.

[0066] If the combination is valid (S14: YES), the hypothesis generating unit 30 determines whether the selected combination is a special case of another hypothesis included in the hypothesis set (S15).

[0067] For example, Figure 4 Combination C05 And combination C06 is by supplying the combination C04 Suppose that the generation unit 30 determines that such combinations C05 and C06 are the combinations of C04. special circumstances.

[0068] If it is a special case ( S15 : Yes), the hypothesis generation unit 30 proceeds to S17 without adding the selected combination to the hypothesis set.

[0069] Figure 5 is an illustration used to illustrate the generation of hypotheses. Figure 5 As shown, the hypothesis generation unit 30 is executed so that as The special case combinations (combinations C05, C06) are omitted and the The combination C04a is taken as the hypothesis set.

[0070] If it is not a special case (S15: No), the hypothesis generation unit 30 adds the selected combination to the hypothesis set of the hypothesis set data 23 (S16). Next, the hypothesis generation unit 30 determines whether all the combinations listed in S12 have been selected (S17). If there is a combination that has not been selected (S17: No), the hypothesis generation unit 30 returns to S13.

[0071] By repeating S13 to S17 , the hypothesis generating unit 30 lists all hypotheses in the hypothesis set that can correctly explain that the prediction is made based on the target variable of the training data 21 .

[0072] Figure 6 is an explanatory diagram for explaining the generation of a hypothesis, and specifically, for explaining using a Carnot diagram example. Figure 4 and Figure 5 A diagram of the contents.

[0073] like Figure 6 As shown, it is assumed that the generation unit 30 generates the following combinations according to A (the remaining three explanatory variables are "unused = *") (S31) and The combinations are changed in the order of combinations (the remaining three explanatory variables are "unused = *")..., and valid combinations are checked (S31 to S35...).

[0074] Here, S33 The combination of corresponds to the training data with the target variable being + (P 1 , P 3 and P4 That is, in S33, the training data (P 1 , P 3 , P 4 ) is equal to or greater than a predetermined value. The combination of is determined as a valid combination classified as + category (hypothesis). Note that in the following processing, Add letters to the combination.

[0075] Next, it is assumed that the generation unit 30 checks all combinations in which three explanatory variables are "unused = *", and then starts checking combinations in which two explanatory variables are "unused = *" (S34). Here, The combination of corresponds to the training data with the target variable being + (P 1 and P 2 That is, in S35, the training data (P 1 and P 2 ) is equal to or greater than a predetermined value. Therefore, The combination of is determined to be a valid combination classified into the + category (hypothesis).

[0076] Figure 7 is an explanatory diagram showing an example of the generated hypothesis. Figure 7 As shown, according to the training data (P 1 To P 4 and N 1 To N 3 ), generates hypotheses whose classification results are + or - among hypotheses H1 to H11, and stores them in the storage unit 20 as hypothesis set data 23.

[0077] Each of the assumptions H1 to H11 is an independent assumption based on the following requirement: correctly describe the training data (P 1 To P 4 and N 1 To N 3 ) is classified as + or -. Therefore, contradictory hypotheses can be included, for example, hypothesis H2 and hypothesis H6.

[0078] In addition, for those not included in the training data (P 1 To P 4 and N 1 To N 3 ) in the input data (IN 1 , IN 2 and IN 3 ), the prediction results can be obtained based on the matching assumptions in assumptions H1 to H11.

[0079] Figure 8 It is used to describe and input data (IN 1 , IN 2 , IN 3 ) is a diagram illustrating the matching hypothesis. Figure 8 As shown, for the input data IN 1 , assuming H2 Hypothesis H6 And hypothesis H8 is a matching assumption. In addition, for the input data IN 2 , assuming that H4 Hypothesis H5 Hypothesis H7 And hypothesis H9 is a matching assumption. In addition, for the input data IN 3 , assuming H1 Hypothesis H7 And hypothesis H8 is the matching assumption.

[0080] Refer to Figure 2 If there is no combination that has not been selected (S17: Yes), the learning unit 40 selects the combination based on whether the combination is selected for the training data (P 1 To P 4 , N 1 To N 3 ) establishes each hypothesis (H1 to H11) included in the hypothesis set of the hypothesis set data 23 to calculate the weight of each hypothesis (H1 to H11). Next, the learning unit 40 stores the calculation result as the weight data 24 in the storage unit 20.

[0081] For example, the weight calculation in the learning unit 40 may be any one of the following three methods.

[0082] - Set the weight of all rules (H1 to H11) to 1 (majority decision based on the number of rules).

[0083] - Based on the training data (P) supporting (corresponding to) the rules (H1 to H11) 1 To P 4 and N 1 To N 3 ) to set the weight.

[0084] -Execution is performed by applying the training data (P 1 To P 4 and N 1 To N 3 ) is weighted by the logistic regression.

[0085] Fig. 9 is a diagram for explaining weighting by logistic regression. In logistic regression, Fig. 9 As shown, the training data (P 1 To P 4 and N 1 To N 3 ) is applied to the model formula to obtain the weights (β 1 To β 11 ). For example, the model formula corresponds to a score function for obtaining a prediction score.

[0086] Here, the learning unit 40 may select a hypothesis according to the weight of each hypothesis ( H1 to H11 ) obtained by logistic regression or the like.

[0087] Fig.10 is an illustration used to illustrate the choice of hypothesis. Fig.10 As shown, the learning unit 40 calculates the weights (β 1 To β 11 ) selects the main hypothesis with a weight equal to or greater than a predetermined value and having a greater impact on the prediction result. In the example shown, the weight is not 0 and is selected and AC’s five hypotheses H1 to H3, H8, and H11 as the main hypotheses.

[0088] Refer to Figure 2 , the operation at the time of prediction (S2) will be described. If S2 starts, the input unit 10 receives input data 22 to be predicted, and stores the received input data 22 in the storage unit 20 (S21).

[0089] Next, the prediction unit 50 obtains data by aggregating each hypothesis listed in the hypothesis set according to a predetermined sequence condition (e.g., processing order) based on the hypothesis set data 23, and generates tree structure data including attribute values ​​(explanatory variable values) as branches (S22).

[0090] Fig.11 This is an illustration for explaining the generation of tree structure data. Fig.11As shown, the hypothesis set data 23 has a set of rules (hypotheses) for explaining the correspondence between variables A, B, C... and labels (P·N), for example, A=1∧B=0→P, A=1∧C=0→N, B=0∧D=0→N... The prediction unit 50 introduces a sequential relationship (A=1→B=1→P) based on the sequential condition (A→B→C→...) for each hypothesis included in the hypothesis set data 23, and thus represents each hypothesis by a path. Next, the prediction unit 50 connects the paths of each hypothesis in order starting from the attribute value close to the root, and aggregates the common parts, thereby generating tree structure data 23a in which the attribute values ​​are converted into a tree structure as branches.

[0091] Next, the prediction unit 50 performs pruning by applying the known attribute values ​​included in the input data 22 to the tree structure data 23a, and then obtains the possibility of obtaining a value of a predetermined label for the unknown attribute value, thereby determining the contribution degree for making the attribute value reach the predetermined label (S23). The prediction unit 50 stores the prediction result as the result data 25 in the storage unit 20 through the determination of the contribution degree.

[0092] FIG. 12A to FIG. 12C : is an explanatory diagram for explaining an example of calculation of the probability of obtaining a value of a predetermined label. Fig. 12A As shown, it is assumed that the tree structure data 23a having attribute values ​​A, B, C, and D as branches is obtained by the hypothesis set data 23 (for example, A=1∧B=0→P, A=1∧C=0→N, B=0∧D=0→N...). In addition, it is also assumed that the input data 22 includes a known action (a part of the explanation variable) of A=1 and B=0 and a target label (P). Therefore, it is assumed that the prediction unit 50 performs pruning using A=1 and B=0 in the tree structure in the tree structure data 23a, and regards the search of the branches A and B as completed (determined).

[0093] like Fig. 12B As shown, in the case where the unknown attribute value C is determined to be C=0, the prediction unit 50 calculates the possibility of obtaining the value of the label P or N by tracing the path of the tree structure in the tree structure data 23a.

[0094] For example, for label P, it is found that the path (rule) of A=1→B=0→P is established when C=0. In addition, for label N, it is found that the path (rule) of A=1→C=0→N is established when C=0. The path (rule) B=0 and D=0 passing through D after C is uncertain, but it is found that the path is likely to be established as label N when C=0.

[0095] The prediction unit 50 obtains the possibility of obtaining the value of the label P or N when C = 0 based on the number of P or N (including those that may also be established) established when C = 0 relative to the total number of paths (rules). In addition, the prediction unit 50 can obtain the possibility of obtaining the value of the label P or N when C = 0 by weighting the score function of the path (rule) where P or N is established.

[0096] In addition, if Fig. 12C As shown, in the case where the unknown attribute value C is determined to be C=1, the prediction unit 50 calculates the possibility of obtaining the value of the label P or N by tracing the path of the tree structure in the tree structure data 23a.

[0097] For example, for label P, it is found that when C = 1, a path (rule) of A = 1 → B = 0 → P is established. Also, for label N, it is found that when C = 1, a path (rule) of B = 0 and D = 0 is likely to be established.

[0098] The prediction unit 50 obtains the possibility of obtaining the value of the label P or N when C = 1 based on the number of P or N (including those that may also be established) established when C = 1 relative to the total number of paths (rules). In addition, the prediction unit 50 can obtain the possibility of obtaining the value of the label P or N when C = 1 by weighting the score function of the path (rule) that establishes P or N.

[0099] Fig.13 is an explanatory diagram showing the contribution degree. Fig.13 As shown, Example 71 illustrates a case where C=0 and the number of rules satisfying P is 1 and the number of rules satisfying N is 1. Furthermore, the number of rules that are likely (likely) to satisfy P is 0, and the number of rules that are likely (likely) to satisfy N is 1. Therefore, the prediction unit 50 calculates the possibility that a value higher than the value of the label P can be obtained when C=0.

[0100] In addition, Example 72 illustrates a case where C=1 and the number of rules satisfying P is 1 and the number of rules satisfying N is 0. Furthermore, the number of rules that are likely (likely) to satisfy P is 0, and the number of rules that are likely (likely) to satisfy N is 1. Therefore, the prediction unit 50 calculates the possibility that a value higher than the value of the label P can be obtained when C=1.

[0101] The prediction unit 50 compares Example 71 and Example 72 (for example, compares the likelihood of P) to obtain C=1 in which the likelihood of P becomes higher as a prediction result. In addition, for example, the prediction unit 50 can determine the contribution of the attribute value C to the value of the label P based on the difference between the likelihood of obtaining the value of the label P when C=0 and the likelihood of obtaining the value of the label P when C=1. For example, the first case is a case where, when C=0, the number of rules that are likely (likely) to satisfy P is 0, and when C=1, the number of rules that are likely (likely) to satisfy P is 1. In addition, the second case is a case where, when C=0, the number of rules that are likely (likely) to satisfy P is 0, and when C=1, the number of rules that are likely (likely) to satisfy P is 2. Comparing the first case with the second case, when C=1, the number of (likely) rules that are likely (likely) to satisfy P increases by 1 in the first case and increases by 2 in the second case, compared to the case when C=0. In this case, the prediction unit 50 may determine that the attribute value of C=1 in the second case contributes more to the value of the label P than the attribute value of C=1 in the first case. Alternatively, the prediction unit 50 may determine that an attribute value having a large difference between the number of rules that are likely (likely) to satisfy P when the attribute values ​​are 0 and 1 among the plurality of attribute values ​​contributes more to the value of the label P.

[0102] FIG. 14A to FIG. 14C It is an explanatory diagram for explaining an example of calculation of the probability of obtaining a value of a predetermined label. FIG. 14A to FIG. 14C The example shows that the variable ratio FIG. 12A to FIG. 12C The variables shown in FIG.

[0103] Specifically, it is assumed that the tree structure data 23a having attribute values ​​A, B, C, D, and E as branches is obtained by the hypothesis set data 23 (for example, A=1∧C=0∧D=1→N, A=1∧C=0∧E=1→P, B=0∧D=0→N...). In addition, it is also assumed that the input data 22 includes a known action (a part of the explanatory variable) of A=1 and B=0 and a target label (P). Therefore, it is assumed that the prediction unit 50 performs pruning using A=1 and B=0 in the tree structure in the tree structure data 23a, and regards the search of the branches A and B as completed (determined).

[0104] like Fig. 14B As shown, in the case where the unknown attribute value C is determined to be C=0, the prediction unit 50 calculates the possibility of obtaining the value of the label P or N by tracing the path of the tree structure in the tree structure data 23a.

[0105] For example, for label P, it is found that when C = 0, the path (rule) of A = 1 → C = 0 → E = 1 is likely to be established. In addition, for label N, it is found that when C = 0, the paths (rules) of A = 1 → C = 0 → E = 0 and B = 0 → D = 0 are likely to be established.

[0106] In addition, if Fig. 14C As shown, in the case where the unknown attribute value C is determined to be C=1, the prediction unit 50 calculates the possibility of obtaining the value of the label P or N by tracing the path of the tree structure in the tree structure data 23a.

[0107] For example, for label P, the path (likely) established when C=1 is 0. In addition, for label N, it is found that when C=1, it is likely to establish the paths A=1→C=1→D=1 and B=0→D=0 (rules).

[0108] Therefore, the prediction unit 50 compares the case of C=0 with the case of C=1 (for example, compares the likelihood of P) to obtain C=0 where the likelihood of P becomes higher as a prediction result.

[0109] Refer to Figure 2 , the output unit 60 reads the result data 25 from the storage unit 20, and outputs the prediction result predicted by the prediction unit 50 through the above processing to a display, a file, etc. (S24). That is, the output unit 60 outputs the action (attribute value) in the unknown attribute (explanatory variable) that becomes the target label after the known action is performed.

[0110] Fig.15 is an explanatory diagram for explaining the output of the prediction results. Fig.15 As shown, the information processing device 1 receives input data 22, which includes intermediate data (for example, known attribute values ​​obtained through a manufacturing process), labels to be satisfied (non-defective products), functions for score calculation, and the like.

[0111] For example, for an unknown action in a subsequent manufacturing process, the prediction unit 50 of the information processing device 1 predicts an optimal action (attribute value) that becomes a label to be satisfied after performing a known action indicated by the midway data. Next, the output unit 60 outputs the optimal action based on the predicted attribute value as the result data 25. Thus, the user can easily confirm, for example, an action that becomes a label to be satisfied (non-defective product) in a subsequent manufacturing process.

[0112] As described above, the information processing device 1 includes an input unit 10 and a prediction unit 50. The input unit 10 receives input data 22 to be predicted. The prediction unit 50 generates tree structure data 23a, which is data obtained by aggregating multiple hypotheses according to predetermined order conditions for multiple attributes based on hypothesis set data 23 including multiple rule information (hypotheses) and has attribute values ​​as branches, and each rule information in the multiple rule information (hypothesis) is indicated by the association between the combination of attribute values ​​of multiple attributes and a label. In addition, when the attribute value of a predetermined attribute among the multiple attributes in the tree structure data 23a is determined as a predetermined value based on the input data 22, the prediction unit 50 determines the contribution of making the predetermined value of the predetermined attribute reach the predetermined label based on the possibility of obtaining the value of the predetermined label as a result of the determination.

[0113] As described above, the information processing device 1 generates the tree structure data 23a having the attribute values ​​as branches according to the predetermined order conditions for the plurality of attributes, and obtains the value of the predetermined attribute using the generated tree structure data 23a. For this reason, in the information processing device 1, prediction can be efficiently performed by trying actions along the tree structure without having to perform all actions (attributes) based on each of the plurality of classification rules (in turn).

[0114] In addition, the prediction unit 50 performs pruning by applying the known attribute values ​​included in the input data 22 to the tree structure data 23. The prediction unit 50 determines the contribution to the fact that, for an attribute whose attribute value is unknown among a plurality of attributes, the determination result when the unknown attribute value of the attribute is determined to be a predetermined value is a label included in the input data 22. Thus, the information processing device 1 can effectively predict the unknown attributes (the remaining explanatory variables) based on the known attribute values ​​(a part of the explanatory variables) and the labels (target variables) included in the input data 22.

[0115] In addition, the prediction unit 50 obtains the contribution to the fact that the determination result when the attribute value of the predetermined attribute is determined to be the first value (for example, C=0) becomes the predetermined label (for example, P). In addition, the prediction unit 50 obtains the contribution to the fact that the determination result when the attribute value of the predetermined attribute is determined to be the second value (for example, C=1) becomes the predetermined label (for example, P). Next, the prediction unit 50 compares the contribution when the attribute value of the predetermined attribute is set to the first value with the contribution when the attribute value of the predetermined attribute is set to the second value, and determines whether the predetermined attribute is the first value or the second value. For example, the information processing device 1 can predict the attribute value with a higher contribution as a result.

[0116] Note that each component of each device shown does not necessarily need to be physically configured as shown. That is, the specific form of distribution / integration of each device is not limited to the form of distribution / integration shown in the drawings, and all or part of it can be functionally or physically distributed / integrated into any unit according to various loads and usage conditions.

[0117] The various processing functions performed by the information processing device 1 can be performed in whole or arbitrarily on a CPU (or a microcomputer, such as an MPU or a microcontroller unit (MCU)). In addition, it goes without saying that the various processing functions can be performed in whole or in part on a program analyzed and executed by the CPU (or a microcomputer, such as an MPU or an MCU) or by hardware of wired logic. In addition, the various processing functions performed by the information processing device 1 can be performed by multiple computers in collaboration with each other through cloud computing.

[0118] However, various processes described in the above embodiments can be realized by having a computer execute a prepared program. Therefore, hereinafter, an example of a computer (hardware) that executes a program having the same functions as the above embodiments will be described. Fig.16 is an explanatory diagram for explaining a hardware configuration example of the information processing device 1 according to the embodiment.

[0119] like Fig.16 As shown, the information processing device 1 includes a CPU 101 that performs various types of arithmetic processing, an input device 102 that receives data input, a monitor 103, and a speaker 104. In addition, the information processing device 1 includes a medium reading device 105 that reads a program or the like from a storage medium, an interface device 106 for connecting to various devices, and a communication device 107 for communicating with an external device in a wired manner or wirelessly. In addition, the information processing device 1 includes a RAM 108 that temporarily stores various information and a hard disk device 109. In addition, each unit 101 to 109 in the information processing device 1 is connected to a bus 110.

[0120] The hard disk device 109 stores a program 111 for executing the various processes described in the above embodiment. In addition, the hard disk device 109 stores various data 112 (e.g., training data 21, input data 22, hypothesis set data 23, weight data 24, and result data 25) referenced by the storage program 111. For example, the input device 102 receives operation information input from an operator of the information processing device 1. For example, the monitor 103 displays various display screens operated by the operator. For example, the interface device 106 is connected to a printing device. The communication device 107 is connected to a communication network (e.g., a local area network (LAN)), and exchanges various information with an external device via the communication network.

[0121] The CPU 101 reads the program 111 stored in the hard disk device 109, expands the program 111 into the RAM 108, and executes the program to perform various processes related to the input unit 10, the hypothesis generation unit 30, the learning unit 40, the prediction unit 50, and the output unit 60. Note that the program 111 may not be stored in the hard disk device 109. For example, the program 111 stored in a storage medium readable by the information processing device 1 can be read and executed by the information processing device 1. For example, the storage medium readable by the information processing device 1 corresponds to a portable recording medium, such as a CD-ROM or DVD disk, a universal serial bus (USB) memory, a semiconductor memory (e.g., a flash memory), a hard disk drive, etc. In addition, the program 111 may be stored in a device connected to a public line, the Internet, a LAN, etc., and the information processing device 1 can read the program from the device and execute the program.

[0122] According to one embodiment of the present invention, prediction of input data can be performed efficiently.

Claims

1. A computer-readable recording medium storing a prediction program for causing a computer to execute a process, the process comprising: receiving input data to be predicted; generating tree structure data by aggregating a plurality of rule information according to a predetermined sequence condition for a plurality of attributes, wherein each of the plurality of rule information is indicated by an association of a combination of attribute values ​​of the plurality of attributes with a label, the tree structure data including attribute values ​​as branches, wherein the predetermined sequence condition indicates a processing sequence in a manufacturing process, and the attribute values ​​of the attributes indicate actions including observation values ​​or control setting values ​​in the manufacturing process; as well as When the attribute value of a predetermined attribute among the multiple attributes in the tree structure data is determined to be a predetermined value based on the input data, based on the possibility of obtaining a value of a predetermined label as a determination result of the predetermined value of the predetermined attribute, the contribution of making the determination result of the predetermined value of the predetermined attribute reach the predetermined label is determined.

2. The computer-readable recording medium according to claim 1, wherein: The determination processing performs pruning by applying known attribute values ​​included in the input data to the tree structure data, and then, for attributes whose attribute values ​​are unknown among the multiple attributes, determines the contribution of the determination result when the unknown attribute value of the attribute is determined to be a predetermined value to the label included in the input data.

3. The computer-readable recording medium according to claim 1 or 2, wherein: The determination processing compares the contribution of the determination result when the attribute value of the predetermined attribute is determined to be a first value to the predetermined tag with the contribution of the determination result when the attribute value of the predetermined attribute is determined to be a second value to the predetermined tag, so as to determine whether the predetermined attribute is the first value or the second value.

4. A prediction method, which causes a computer to perform the following processing: receiving input data to be predicted; Tree structure data is generated by aggregating a plurality of rule information according to a predetermined sequence condition for a plurality of attributes, wherein each of the plurality of rule information is indicated by an association between a combination of attribute values ​​of the plurality of attributes and a label, the tree structure data including attribute values ​​as branches, wherein, The predetermined sequence condition indicates a process sequence in a manufacturing process, and the attribute value of the attribute indicates an action including an observation value or a control setting value in the manufacturing process; as well as When the attribute value of a predetermined attribute among the multiple attributes in the tree structure data is determined to be a predetermined value based on the input data, based on the possibility of obtaining a value of a predetermined label as a determination result of the predetermined value of the predetermined attribute, the contribution of making the determination result of the predetermined value of the predetermined attribute reach the predetermined label is determined.

5. The prediction method according to claim 4, wherein: The determination processing performs pruning by applying known attribute values ​​included in the input data to the tree structure data, and then, for attributes whose attribute values ​​are unknown among the multiple attributes, determines the contribution of the determination result when the unknown attribute value of the attribute is determined to be a predetermined value to the label included in the input data.

6. The prediction method according to claim 4 or 5, wherein: The determination processing compares the contribution of the determination result when the attribute value of the predetermined attribute is determined to be a first value to the predetermined tag with the contribution of the determination result when the attribute value of the predetermined attribute is determined to be a second value to the predetermined tag, so as to determine whether the predetermined attribute is the first value or the second value.

7. A prediction device, comprising: an input unit that receives input data to be predicted; as well as A prediction unit, which generates tree structure data by aggregating multiple rule information according to predetermined sequence conditions for multiple attributes, wherein each of the multiple rule information is indicated by the association between the combination of attribute values ​​of the multiple attributes and a label, and the tree structure data is obtained by aggregating the multiple rule information, and the tree structure data includes attribute values ​​as branches, wherein the predetermined sequence conditions indicate the processing sequence in the manufacturing process, and the attribute value indication of the attribute includes an observation value or a control setting value in the manufacturing process; and when the attribute value of a predetermined attribute among the multiple attributes in the tree structure data is determined as a predetermined value based on the input data, based on the possibility of obtaining a value of a predetermined label as a determination result of the predetermined value of the predetermined attribute, determine the contribution of making the determination result of the predetermined value of the predetermined attribute reach the predetermined label.

8. The prediction device according to claim 7, wherein: The prediction unit performs pruning by applying known attribute values ​​included in the input data to the tree structure data, and then determines, for attributes whose attribute values ​​are unknown among the multiple attributes, a contribution that makes a determination result when the unknown attribute value of the attribute is determined to be a predetermined value reach a label included in the input data.

9. The prediction device according to claim 7 or 8, wherein: The prediction unit compares a contribution degree that causes a determination result when the attribute value of the predetermined attribute is determined to be a first value to reach the predetermined label with a contribution degree that causes a determination result when the attribute value of the predetermined attribute is determined to be a second value to reach the predetermined label, to determine whether the predetermined attribute is the first value or the second value.