Information processing device, information processing method, and program

Abstract

An information processing device 100 according to the present disclosure comprises: a candidate rule generation unit 121 that generates a candidate rule set that is a set of conditional prediction rules on the basis of a set of training examples; and a search unit 122 that searches for and generates, from among combinations of prediction rules selected from the candidate rule set, a solution rule satisfying a predetermined criterion of prediction performance on the basis of prediction performances for the training examples. The candidate rule generation unit 121 generates a new prediction rule to be included in the candidate rule set on the basis of the solution rule.

Description

Information processing device, information processing method, and program

[0001] The present disclosure relates to an information processing device, an information processing method, and a program.

[0002] Predictions for input data are made using machine learning models in a variety of fields. Examples of machine learning models include rule-based models such as those described in Patent Literature 1, which use a training case set to learn a combination of rules that will become a solution from a candidate rule set.

[0003] International Publication No. 2022 / 044221

[0004] However, the number of rules used in the rule-based model is enormous, making it difficult to find an appropriate rule, which poses a problem that it is not possible to further improve the prediction accuracy of the rule-based model.

[0005] An object of the present disclosure is to solve the above-mentioned problem that it is not possible to further improve the prediction accuracy in a rule-based model.

[0006] An information processing device according to one aspect of the present disclosure includes: a candidate rule generation unit that generates a candidate rule set, which is a set of conditional prediction rules, based on a set of training cases; and a search unit that searches for and generates a solution rule, from combinations of the prediction rules selected from the candidate rule set, whose prediction performance satisfies a preset criterion based on their prediction performance on the training cases, wherein the candidate rule generation unit generates a new prediction rule to be included in the candidate rule set based on the solution rule.An information processing method according to one aspect of the present disclosure includes: generating a candidate rule set, which is a set of conditional prediction rules, based on a set of training cases; searching for and generating a solution rule, from combinations of the prediction rules selected from the candidate rule set, whose prediction performance satisfies a preset criterion based on their prediction performance on the training cases, and generating the new prediction rule to be included in the candidate rule set based on the solution rule. Furthermore, a program according to one aspect of the present disclosure has a configuration that causes a computer to execute the following processes: generate a candidate rule set, which is a set of conditional prediction rules, based on a set of training cases; search for and generate a solution rule, from among combinations of prediction rules selected from the candidate rule set, whose prediction performance satisfies a preset standard based on prediction performance for the training cases; and generate a new prediction rule to be included in the candidate rule set, based on the solution rule.

[0007] With the above-described configuration, the present disclosure can further improve the prediction accuracy in a rule-based model.

[0008] FIG. 1 is a block diagram showing a configuration of an information processing device according to the present disclosure. FIG. 2 is a diagram showing a processing state by an information processing device according to the present disclosure. FIG. 3 is a flowchart showing a processing operation by an information processing device according to the present disclosure. FIG. 4 is a diagram showing a processing state by an information processing device according to the present disclosure. FIG. 5 is a block diagram showing a hardware configuration of an information processing device according to the present disclosure. FIG. 6 is a block diagram showing a configuration of an information processing device according to the present disclosure.

[0009] First Embodiment A first embodiment of the present disclosure will be described with reference to the drawings. Note that the drawings may be relevant to any embodiment.

[0010] The information processing device 10 in this embodiment generates a rule-based model using training cases. In particular, in this embodiment, as shown in Fig. 2, the information processing device 10 generates a candidate rule set R, which is a set of candidate prediction rules, from an entire rule set D of conditional prediction rules, and further searches for a solution rule set S consisting of a combination of prediction rules included in the candidate rule set R, thereby generating a rule-based model consisting of the solution rule set S. The configuration and operation of the information processing device 10 will be described in detail below.

[0011] The information processing device 10 is composed of one or more information processing devices each including a calculation device and a storage device. As shown in FIG. 1 , the information processing device 10 includes an input unit 11, a candidate rule generation unit 12, and a solution rule search unit 13. The functions of the input unit 11, the candidate rule generation unit 12, and the solution rule search unit 13 can be realized by the calculation device executing a program for realizing each function stored in the storage device. The information processing device 10 also includes a training example storage unit 14, a candidate rule storage unit 15, and a solution rule storage unit 16. The training example storage unit 14, the candidate rule storage unit 15, and the solution rule storage unit 16 are each composed of a storage device. The functions and operations of each component are described below.

[0012] The input unit 11 receives input of a training example set D that the information processing device 10 uses to train the rule-based model, and stores it in the training example storage unit 14 (step S1 in FIG. 3). Here, the training example set is defined as D={(x １ , y １ ),...,(x ｎ , y ｎ )}, and for 1≦i≦n, x ｉ Enter y ｉ is called a label. ｉis given as a real number or an element of a label set. In this embodiment, when the training example set D as described above is given, a regression problem or classification problem, which is a problem of learning a model f: X → Y, is considered, and the information processing device 10 derives a model f consisting of a set S of prediction rules from the training example set D, as will be described later. ｓ For example, the information processing device 10 learns a model f that is a set S of prediction rules whose prediction performance score(f, D) for a training example set D satisfies a preset standard. ｓ In a regression problem, the prediction performance score(f, D) is the mean squared error of the predicted value of the model f for the training example set D, or the value of an objective function obtained by adding an L1 regularization term to the mean squared error. In this case, the smaller the value of the objective function, the higher the prediction performance. In a classification problem, the prediction performance score(f, D) is the accuracy or AUC (Area Under the ROC Curve) score of the model f for the training example set D. In this case, the higher the accuracy or AUC score, the higher the prediction performance.

[0013] First, in phase t=1 of the solution search shown in FIG. 2 (step S2 in FIG. 3), the candidate rule generation unit 12 generates a candidate rule set R, which is a set of prediction rules, based on the training example set D. １ and stores it in the candidate rule storage unit 15 (step S3 in FIG. 3). At this time, the candidate rule generation unit 12 generates a candidate rule set R, which is a set of multiple prediction rules, from a set of all possible conditional prediction rules (all rule sets) from the training case set D using a known method such as a random forest or rule mining technique. ｔ In the example of FIG. 2, the candidate rule set R １ As a result, a set of prediction rule candidates indicated by a plurality of white circles is generated.

[0014] The solution rule search unit 13 (search unit) searches for a candidate rule set R ｔ, and searches for a combination of prediction rules that satisfies a preset criterion, such as high prediction performance for a training case set, among the combinations of prediction rules by machine learning (step S4 in FIG. 3). ｔ is generated and stored in the solution rule storage unit 16. In the example of FIG. 2, in the solution search phase t=1, the candidate rule set R １ From the solution rule set S １ is searched for and generated.

[0015] In this embodiment, as indicated by the arrow L in FIG. 2, the candidate rule generating unit 12 further generates a solution rule set S ｔ Based on the new candidate rule set R ｔ＋１ and stores it in the candidate rule storage unit 15. Also, the solution rule search unit 13 generates a new candidate rule set R ｔ＋１ Based on this, a combination of prediction rules is selected in the same manner as above, and a new solution rule set S ｔ＋１ The candidate rule generating unit 12 and the solution rule searching unit 13 search for and generate the candidate rule set R ｔ and the solution rule set S ｔ For example, in the example of FIG. 2, in the subsequent solution search phase t=2 (No in step S5 of FIG. 3, S6), the solution rule set S generated in the solution search phase t=1 is generated. １ Based on the new candidate rule set R ２ (step S7 in FIG. 3), and a new candidate rule set R ２ From this, a new solution rule set S ２ (Step S4 in FIG. 3). The above process is then repeated until the solution search phase t=T (Step S5 in FIG. 3), and a new candidate rule set and a new solution rule set are generated in each solution search phase.

[0016] Here, the details of the processing from the second time onwards, that is, from solution search phase t=2 onwards, will be described. The candidate rule generating unit 12 generates a candidate rule set R ｔBy adding and including new prediction rules to the rule set R ｔ＋１ The new prediction rule to be added is generated as the solution rule set S generated in the previous solution search phase t. ｔ At this time, the candidate rule generation unit 12 generates the solution rule set S ｔ The new prediction rule is generated so that the new prediction rule that further improves the prediction performance of the new candidate rule generation unit 12 can be added to the new candidate rule generation unit 12.

[0017] The new candidate rule set R ｔ＋１ First, in a first example, the candidate rule generation unit 12 generates a new candidate rule set R as follows by a method using a random forest. ｔ＋１ Generate.

[0018] In the first example, the candidate rule generating unit 12 first selects the solution rule set S generated in the previous solution search phase t. ｔ Depending on the prediction performance for the training example set D by the ｔ＋１ Here, as a predetermined process for the training example set D, the candidate rule generation unit 12 generates a solution rule set S ｔ Then, the candidate rule generation unit 12 extracts new prediction rules from the remaining training example set D using a random forest, and generates a new candidate rule set R including the extracted new prediction rules. ｔ＋１ The candidate rule generating unit 12 generates the solution rule set S ｔ The solution rule set S is not limited to removing the training examples correctly classified by ｔTraining cases that are judged to have high predictive performance, such as when predictions made by the random forest satisfy a predetermined standard, may be removed from the training case set D. In addition, the candidate rule generation unit 12 may generate a new candidate rule set from the remaining training case set D using a method other than the random forest. ｔ＋１ may be generated.

[0019] In the first example, the candidate rule generating unit 12 performs a predetermined process on the training case set D to generate a solution rule set S for each training case. ｔ For example, the candidate rule generating unit 12 may perform a process of assigning weights according to the prediction performance of the solution rule set S ｔ The training cases that are correctly classified by the rule generator 12 may be assigned a weight that is smaller than that of the other training cases, and the training case set D may be updated. In other words, the candidate rule generator 12 assigns a solution rule set S ｔ The training example set D may be updated by assigning weights that decrease as the prediction performance by the training example set D increases. Then, the candidate rule generation unit 12 extracts new prediction rules using a random forest from the training example set D updated with the weights, and generates a new candidate rule set R including the extracted new prediction rules. ｔ＋１ The candidate rule generating unit 12 may generate a new candidate rule set R from the updated training example set D using a method other than the random forest. ｔ＋１ may be generated.

[0020] Next, the candidate rule generating unit 12 generates a new candidate rule set R ｔ＋１ In the second example, the candidate rule generating unit 12 generates a new candidate rule set R as follows by a method using a random search. ｔ＋１ In the second example, the candidate rule generating unit 12 generates a new candidate rule set R ｔ When generating the prediction rule, as will be described below, the process of adoption phase t' for adopting the prediction rule to be added is repeated from t'=1 to t'=T'.

[0021] First, in adoption phase t′=1, the candidate rule generation unit 12 calculates the candidate rule set R ｔ The candidate rule generating unit 12 may randomly select a prediction rule from the solution rule set S ｔ In the example of FIG. 4, a prediction rule may be selected that improves prediction performance when included in the candidate rule set R. ｔ The prediction rule of rule 2 "x3<5.0 AND x4>1.0 AND x6<2.0" is randomly selected from the above and adopted as the adopted rule r.

[0022] Next, the candidate rule generation unit 12 applies random perturbation to the adopted rule r to generate a modified adopted rule r' (modified prediction rule), which is a new prediction rule obtained by modifying the content of the adopted rule r. In the example of Fig. 4, the modified adopted rule r' is generated by modifying part of the content of the adopted rule r, "x4 > 1.0", to "x4 > 1.5".

[0023] Then, the candidate rule generating unit 12 generates the solution rule set S ｔ , the prediction performance Score (f Ｓｔ∪ｒ ) and the prediction performance Score (f Ｓｔ∪ｒ' ) and are calculated. In the example of FIG. ｔ When the adopted rule r is included, the prediction performance score (f Ｓｔ∪ｒ ) is "0.8", and the solution rule set S ｔ The prediction performance score (f Ｓｔ∪ｒ' ) is "0.9". Then, the candidate rule generating unit 12 selects the solution rule set S ｔThe rule included in is adopted as the new adopted rule r. In the example of Figure 4, the prediction performance when the modified adopted rule r' is included is higher than the prediction performance Score when the adopted rule r is included, so the modified adopted rule r' is adopted as the new adopted rule r. Note that if the prediction performance Score when the adopted rule r is included is higher than the prediction performance when the modified adopted rule r' is included, the adopted rule r is adopted as is.

[0024] In this way, in the example shown in FIG. 4, when the adoption phase t′=1 is completed, the candidate rule set R ｔ The content of the adopted rule r adopted from the previous step is changed to obtain a modified adopted rule r', which is then adopted as the new adopted rule r.

[0025] The candidate rule generation unit 12 then performs the same processing as above in the subsequent adoption phase t'=2. Specifically, the candidate rule generation unit 12 first applies a random perturbation to the adopted rule r to generate a modified adopted rule r', which is a new prediction rule obtained by modifying the content of the adopted rule r. At this time, the candidate rule generation unit 12 modifies the adopted rule r differently from the modifications made in the previous adoption phase t' to generate the modified adopted rule r'. In the example of FIG. 4, the modified adopted rule r' is generated by modifying part of the content of the adopted rule r, "x3<5.0", to "x3<4.0".

[0026] Then, the candidate rule generating unit 12 generates a solution rule set S ｔ , the prediction performance Score (f Ｓｔ∪ｒ ) and the prediction performance Score (f Ｓｔ∪ｒ' ) and are calculated. In the example of FIG. ｔ When the adopted rule r is included, the prediction performance score (f Ｓｔ∪ｒ ) is "0.9", and the solution rule set S ｔ The prediction performance score (f Ｓｔ∪ｒ') is "0.85". Therefore, the candidate rule generating unit 12 selects the solution rule set S ｔ In other words, in adoption phase t'=2, the modified adopted rule r' whose content has been changed is not adopted.

[0027] The candidate rule generating unit 12 repeatedly performs the above-described adoption process for the adopted rule r until adoption phase t'=T'. Then, the candidate rule generating unit 12 adds the finally adopted adopted rule r to the candidate rule set R of the previous solution search phase t. ｔ and add a new candidate rule set R ｔ＋１ In the example of FIG. 4, a new candidate rule set R is generated by adding the finally adopted rule r "x3<5.0 AND x4>1.5 AND x6<2.0". ｔ＋１ will be generated.

[0028] In this way, the candidate rule generating unit 12 generates a plurality of modified adopted rules r' each having different content, and classifies these into a solution rule set S ｔ The modified adopted rule r′ included in the case where the predictive performance is the highest among the predictive performances when included in the new candidate rule set R ｔ＋１ However, the candidate rule generation unit 12 is not limited to selecting the modified adopted rule r' with the highest predictive performance, and may include multiple modified adopted rules r' that meet a predetermined criterion, such as descending order of predictive performance, in the new candidate rule set. Furthermore, the candidate rule generation unit 12 is not limited to performing the process of adoption phase t' multiple times as described above, but may perform adoption phase t' once and include the generated modified adopted rule r' in the new candidate rule set.

[0029] Then, in each solution search phase t, the solution rule search unit 13 calculates the new candidate rule set R generated as described above. ｔ From the new solution rule set S ｔ At this time, the solution rule search unit 13 searches for and generates a new candidate rule set R ｔ＋１ The prediction rules added to are added to the existing solution rule set S in the previous solution search phase t. ｔThe additional solution rule set S' added to ｔ and generate an existing solution rule set S ｔ and the additional solution rule set S' ｔ Then, the solution rule search unit 13 calculates the prediction performance for the training example set D by using the existing solution rule set S ｔ For the prediction performance by ｔ When a predetermined standard is met, such as when the prediction performance by ｔ The new solution rule set S ｔ＋１ Furthermore, at this time, the solution rule search unit 13 generates a new candidate rule set R for each solution search phase t with respect to the solution rule set in the previous solution search phase. ｔ , and one or more of the newly added prediction rules are added to the solution rule set S ｔ In this way, the solution rule search unit 13 can also apply the greedy learning method.

[0030] In addition, the solution rule set S ｔ When searching for a set of training examples D, the objective function used to evaluate the prediction performance for the set of training examples D is assumed to be submodular with respect to the addition of prediction rules. In other words, the set function used as the objective function has the property that the smaller the set, the greater the change in the function value when a new element is added.

[0031] Here, let us say that function f is a set function whose base is a finite set V = {1,...,n} consisting of n elements, and that f is a function that assigns a real number f(S) to any subset S⊆V of V. In this case, for any S, T⊆V that satisfies S⊆T and any element i∈V that is not included in T, when f(S∪i)-f(S)≧f(T∪i)-f(T) holds, the set function f can be said to be a submodular function.

[0032] As described above, in the information processing device 10 of this embodiment, the generated solution rule set S ｔ Based on the new candidate rule set R ｔ＋１ and generate a new solution rule set Sｔ＋１ As a result, for example, as shown in the left diagram of FIG. 5, when the true decision boundary in the problem is the curve indicated by the solid line, the gray area is the solution rule set S ｔ If the area is predictable in the solution rule set S ｔ＋１ In other words, in this embodiment, the solution rule set S ｔ A new prediction rule not included in the new candidate rule set R ｔ＋１ , so the existing solution rule set S ｔ This allows us to search for rules near the decision boundary that are not covered by the previous method, thereby further improving the prediction accuracy of the rule-based model.

[0033] Second Embodiment Next, a second embodiment of the present disclosure will be described with reference to the drawings. This embodiment shows an outline of the configuration of the information processing device described in the above embodiment. Note that Figures 6 and 7 are diagrams for explaining the configuration, and these drawings may be relevant to any of the embodiments.

[0034] First, the hardware configuration of the information processing device 100 will be described with reference to Fig. 6. The information processing device 100 is configured as a general information processing device, and is equipped with the following hardware configuration, for example: CPU (Central Processing Unit) 101 (arithmetic unit); ROM (Read Only Memory) 102 (storage device); RAM (Random Access Memory) 103 (storage device); programs 104 loaded into RAM 103; storage device 105 storing programs 104; drive device 106 for reading and writing data from and to a storage medium 110 external to the information processing device; communication interface 107 for connecting to a communication network 111 external to the information processing device; input / output interface 108 for inputting and outputting data; and bus 109 for connecting the various components.

[0035] 6 shows an example of the hardware configuration of the information processing device 100, and the hardware configuration of the information processing device is not limited to the above-described case. For example, the information processing device may be configured with only a part of the above-described configuration, such as excluding the drive device 106. Furthermore, the information processing device may use a GPU (Graphics Processing Unit), a DSP (Digital Signal Processor), an MPU (Micro Processing Unit), an FPU (Floating Point Number Processing Unit), a PPU (Physics Processing Unit), a TPU (Tensor Processing Unit), a quantum processor, a microcontroller, or a combination thereof, instead of the above-described CPU.

[0036] Then, by having CPU 101 acquire and execute program group 104, information processing apparatus 100 can construct and be equipped with candidate rule generation unit 121 and search unit 122 shown in FIG. 7. Program group 104 is stored in advance in storage device 105 or ROM 102, for example, and is loaded into RAM 103 and executed by CPU 101 as needed. Program group 104 may be supplied to CPU 101 via communication network 111, or may be stored in advance in storage medium 110, with drive device 106 reading out the programs and supplying them to CPU 101. However, candidate rule generation unit 121 and search unit 122 described above may be constructed using dedicated electronic circuits for realizing such means.

[0037] The candidate rule generation unit 121 generates a candidate rule set, which is a set of conditional prediction rules, based on a set of training cases. The search unit 122 searches for and generates a solution rule, from among combinations of prediction rules selected from the candidate rule set, whose prediction performance satisfies a preset standard based on prediction performance for training cases. Furthermore, the candidate rule generation unit 121 generates a new prediction rule to be included in the candidate rule set based on the solution rule.

[0038] With the above configuration, the present disclosure generates new candidate rules based on the generated solution rules, and then generates new solution rules. This makes it possible to search for rules near decision boundaries that are not covered by existing solution rules. As a result, it is possible to further improve the prediction accuracy of the rule-based model.

[0039] In addition, at least one of the functions of the candidate rule generation unit 121 and the search unit 122 described above may be executed by an information processing device installed and connected anywhere on the network, that is, they may be executed by so-called cloud computing.

[0040] The above-described program can be stored and supplied to a computer using various types of non-transitory computer-readable media. Non-transitory computer-readable media include various types of tangible storage media. Examples of non-transitory computer-readable media include magnetic recording media (e.g., flexible disks, magnetic tapes, hard disk drives), magneto-optical recording media (e.g., magneto-optical disks), CD-ROMs (Read Only Memory), CD-Rs, CD-RWs, and semiconductor memories (e.g., mask ROMs, PROMs (Programmable ROMs), EPROMs (Erasable PROMs), flash ROMs, and RAMs (Random Access Memory)). The program can also be supplied to a computer by various types of transitory computer-readable media. Examples of transitory computer-readable media include electrical signals, optical signals, and electromagnetic waves. The transitory computer-readable media can be supplied to a computer via wired communication paths such as electric wires and optical fibers, or via wireless communication paths.

[0041] Although the present disclosure has been described above with reference to the above-described embodiments, the present disclosure is not limited to the above-described embodiments. Various modifications that can be understood by those skilled in the art can be made to the configuration and details of the present disclosure within the scope of the present disclosure. Furthermore, each of the above-described embodiments can be combined with other embodiments as appropriate.

[0042] <Supplementary Notes> Some or all of the above embodiments can also be described as in the following supplementary notes. Below, an outline of the configurations of an information processing device, an information processing method, and a program according to the present disclosure will be described. However, the present disclosure is not limited to the following configurations. (Supplementary Note 1) An information processing device comprising: a candidate rule generation unit that generates a candidate rule set, which is a set of conditional prediction rules, based on a set of training cases; and a search unit that searches for and generates a solution rule, from combinations of the prediction rules selected from the candidate rule set, whose prediction performance satisfies a predetermined criterion based on prediction performance on the training cases, wherein the candidate rule generation unit generates a new prediction rule to be included in the candidate rule set based on the solution rule. (Supplementary Note 2) The information processing device according to Supplementary Note 1, wherein the candidate rule generation unit generates the new prediction rule to be included in the candidate rule set based on prediction performance of the solution rule on the training cases. (Supplementary Note 3) The information processing device according to Supplementary Note 2, wherein the candidate rule generation unit performs predetermined processing on the training cases in accordance with prediction performance of the solution rule for the training cases, and generates the new prediction rule to be included in the candidate rule set based on the training cases on which the predetermined processing has been performed. (Supplementary Note 4) The information processing device according to Supplementary Note 3, wherein the candidate rule generation unit assigns a weight to the training case in accordance with prediction performance of the solution rule for the training cases, and generates the new prediction rule to be included in the candidate rule set based on the training cases to which the weight has been assigned. (Supplementary Note 5) The information processing device according to Supplementary Note 3, wherein the candidate rule generation unit generates the new prediction rule to be included in the candidate rule set based on the training cases other than the training cases for which the prediction performance of the solution rule for the training cases satisfies a predetermined criterion.(Supplementary Note 6) The information processing device according to Supplementary Note 2, wherein the candidate rule generation unit generates the new prediction rule to be included in the candidate rule set based on prediction performance for the training cases by a first solution rule including a predetermined prediction rule included in the candidate rule set, and prediction performance for the training cases by a second solution rule including a modified prediction rule obtained by modifying the content of the predetermined prediction rule. (Supplementary Note 7) The information processing device according to Supplementary Note 6, wherein the candidate rule generation unit generates the modified prediction rule as the new prediction rule to be included in the candidate rule set when prediction performance for the training cases by the second solution rule is higher than prediction performance for the training cases by the first solution rule. (Supplementary Note 8) The information processing device according to Supplementary Note 7, wherein the candidate rule generation unit generates, as the new prediction rule to be included in the candidate rule set, the changed prediction rule included in the second solution rule having the highest prediction performance among prediction performances on the training cases by a plurality of second solution rules, each including a plurality of changed prediction rules obtained by changing the content of the predetermined prediction rule so that each rule has a different content. (Supplementary Note 9) The information processing device according to Supplementary Note 1, wherein the search unit searches for and generates the new solution rule based on prediction performance on the training cases by a combination of prediction rules obtained by adding the prediction rule newly included in the candidate rule set to an existing solution rule. (Supplementary Note 10) The information processing device according to Supplementary Note 9, wherein the search unit includes the prediction rule added to the existing solution rule in the new solution rule when prediction performance on the training cases by a combination of prediction rules obtained by adding the prediction rule newly included in the candidate rule set to the existing solution rule is higher than prediction performance on the training cases by the existing solution rule. (Supplementary Note 11) The information processing device according to Supplementary Note 9, wherein an objective function used when the search unit evaluates prediction performance for the training examples by the combination of the prediction rules has submodularity with respect to the addition of the prediction rules.(Supplementary Note 12) An information processing method comprising: generating a candidate rule set which is a set of conditional prediction rules based on a set of training cases; searching for and generating a solution rule, from combinations of the prediction rules selected from the candidate rule set, whose prediction performance satisfies a predetermined criterion based on prediction performance on the training cases; and generating a new prediction rule to be included in the candidate rule set based on the solution rule. (Supplementary Note 13) An information processing method according to Supplementary Note 12, wherein the new prediction rule to be included in the candidate rule set is generated based on prediction performance on the training cases by the solution rule. (Supplementary Note 14) An information processing method according to Supplementary Note 13, wherein the training cases are subjected to predetermined processing in accordance with prediction performance on the training cases by the solution rule, and the new prediction rule to be included in the candidate rule set is generated based on the training cases on which the predetermined processing has been performed. (Supplementary Note 15) The information processing method according to Supplementary Note 13, wherein a new prediction rule to be included in the candidate rule set is generated based on prediction performance for the training examples by a first solution rule including a predetermined prediction rule included in the candidate rule set, and prediction performance for the training examples by a second solution rule including a modified prediction rule obtained by modifying the content of the predetermined prediction rule. (Supplementary Note 16) The information processing method according to Supplementary Note 12, wherein a new solution rule is searched for and generated based on prediction performance for the training examples by a combination of prediction rules obtained by adding the prediction rule newly included in the candidate rule set to an existing solution rule. (Supplementary Note 17) A computer-readable storage medium storing a program that causes a computer to execute the following processes: generating a candidate rule set, which is a set of conditional prediction rules, based on a set of training cases; searching for and generating a solution rule, from among combinations of the prediction rules selected from the candidate rule set, whose prediction performance satisfies a predetermined standard based on prediction performance for the training cases; and further generating a new prediction rule to be included in the candidate rule set, based on the solution rule.

[0043] REFERENCE SIGNS LIST 10 Information processing device 11 Input unit 12 Candidate rule generation unit 13 Solution rule search unit 14 Training case storage unit 15 Candidate rule storage unit 16 Solution rule storage unit 100 Information processing device 101 CPU 102 ROM 103 RAM 104 Program group 105 Storage device 106 Drive device 107 Communication interface 108 Input / output interface 109 Bus 110 Storage medium 111 Communication network 121 Candidate rule generation unit 122 Search unit

Claims

1. An information processing device comprising: a candidate rule generation unit that generates a candidate rule set, which is a set of conditional prediction rules, based on a set of training cases; and a search unit that searches for and generates a solution rule, from among combinations of prediction rules selected from the candidate rule set, whose prediction performance satisfies a preset standard based on prediction performance for the training cases, wherein the candidate rule generation unit generates a new prediction rule to be included in the candidate rule set based on the solution rule.

2. An information processing device according to claim 1, wherein the candidate rule generation unit generates a new prediction rule to be included in the candidate rule set based on prediction performance of the solution rule for the training examples.

3. An information processing device according to claim 2, wherein the candidate rule generation unit performs predetermined processing on the training cases according to the predictive performance of the solution rule for the training cases, and generates a new prediction rule to be included in the candidate rule set based on the training cases that have been subjected to the predetermined processing.

4. An information processing device according to claim 3, wherein the candidate rule generation unit assigns a weight to the training case according to the predictive performance of the solution rule for the training case, and generates a new prediction rule to be included in the candidate rule set based on the training case to which the weight has been assigned.

5. An information processing device according to claim 3, wherein the candidate rule generation unit generates new prediction rules to be included in the candidate rule set based on training cases other than those for which the predictive performance of the solution rule for the training cases satisfies a predetermined standard.

6. An information processing device according to claim 2, wherein the candidate rule generation unit generates a new prediction rule to be included in the candidate rule set based on the prediction performance for the training examples by a first solution rule including a predetermined prediction rule included in the candidate rule set, and the prediction performance for the training examples by a second solution rule including a modified prediction rule in which the content of the predetermined prediction rule has been modified.

7. An information processing device according to claim 6, wherein the candidate rule generation unit generates the modified prediction rule as a new prediction rule to be included in the candidate rule set when the prediction performance of the second solution rule for the training examples is higher than the prediction performance of the first solution rule for the training examples.

8. An information processing device according to claim 7, wherein the candidate rule generation unit generates, as a new prediction rule to be included in the candidate rule set, the modified prediction rule included in the second solution rule with the highest predictive performance among the predictive performances for the training cases by the second solution rules each including a plurality of modified prediction rules in which the content of the predetermined prediction rule has been modified to be different from each other.

9. An information processing device according to claim 1, wherein the search unit searches for and generates a new solution rule based on the prediction performance for the training examples obtained by combining the prediction rule obtained by adding the prediction rule newly included in the candidate rule set to an existing solution rule.

10. An information processing device according to claim 9, wherein the search unit, when the prediction performance for the training examples by the combination of the prediction rules obtained by adding the prediction rule newly included in the candidate rule set to the existing solution rule is higher than the prediction performance for the training examples by the existing solution rule, newly includes the prediction rule added to the existing solution rule in the solution rule.

11. An information processing device according to claim 9, wherein the objective function used by the search unit when evaluating the prediction performance for the training examples by the combination of the prediction rules has submodularity with respect to the addition of the prediction rules.

12. An information processing method comprising: generating a candidate rule set, which is a set of conditional prediction rules, based on a set of training cases; searching for and generating a solution rule, from among combinations of prediction rules selected from the candidate rule set, whose prediction performance satisfies a predetermined standard based on the prediction performance for the training cases; and generating a new prediction rule to be included in the candidate rule set based on the solution rule.

13. An information processing method according to claim 12, further comprising generating a new prediction rule to be included in the candidate rule set based on the prediction performance of the solution rule for the training examples.

14. An information processing method according to claim 13, further comprising: performing a predetermined process on the training cases according to the predictive performance of the solution rule for the training cases; and generating a new prediction rule to be included in the candidate rule set based on the training cases that have been subjected to the predetermined process.

15. An information processing method according to claim 13, wherein a new prediction rule to be included in the candidate rule set is generated based on the prediction performance for the training examples by a first solution rule including a predetermined prediction rule included in the candidate rule set, and the prediction performance for the training examples by a second solution rule including a modified prediction rule in which the content of the predetermined prediction rule has been modified.

16. An information processing method according to claim 12, wherein a new solution rule is searched for and generated based on the prediction performance for the training examples obtained by combining the prediction rule newly included in the candidate rule set with an existing solution rule.

17. A computer-readable storage medium storing a program that causes a computer to execute the following processes: generating a candidate rule set, which is a set of conditional prediction rules, based on a set of training cases; searching for and generating a solution rule, from among combinations of prediction rules selected from the candidate rule set, whose prediction performance satisfies a predetermined standard based on the prediction performance for the training cases; and further generating a new prediction rule to be included in the candidate rule set based on the solution rule.

Images