Rule generation apparatus, rule generation method, and non-transitory computer readable medium storing control program

A rule generation apparatus comprising:

• at least one memory storing instructions, and
• at least one processor configured to execute, according to the instructions, a process comprising:
• acquiring a training set including a plurality of examples, wherein the plurality of examples include at least one real example and at least one synthetic example, each of the real example and the synthetic example includes a feature value vector of which vector elements are one or more feature values corresponding to feature parameters different from each other, and a target value, and the feature value and the target value included in each of the real examples are measured values, while each of the synthetic examples is an example formed based on the real example;
• and
• generating a rule group for dividing the training set into a plurality of clusters related to the target values using a rule base model so that a first constraint is satisfied, wherein each of rules in the rule group includes a condition obtained from a condition set including at least one simple condition related to the feature parameters and a predicted value related to the target values,
• wherein the first constraint includes a constraint that each cluster includes at least N (N is a natural number) real examples,
• wherein the generating the rule group comprises:
• generating a condition for each of the rules;
• and
• generating a predicted value of each of the rules, and
• wherein
• the generating the condition further comprises:
• selecting one of a plurality of explanatory variables respectively corresponding to a plurality of the feature parameters of examples included in a division processing target set and generating a plurality of thresholds using the selected explanatory variable, thereby generating a plurality of simple condition candidates for the selected explanatory variable;
• and
• selecting one simple condition candidate from among the plurality of generated simple condition candidates as a use simple condition candidate to be used for processing, forming a first subset and a second subset by dividing a plurality of the examples included in the division processing target set based on the use simple condition candidate, and outputting, for each of the simple condition candidates, an information set including a first information group in which the use simple condition candidate is associated with the first subset as a condition set and a second information group in which a negation of the use simple condition candidate is associated with the second subset as a condition set and then storing the information sets in a first storage,
• the first division processing target set is the training set, and
• the generating the condition further comprises:
• selecting, from a plurality of the stored information sets, the information set in which a first statistical value calculated from the first and the second subsets is the smallest statistical value;
• making a second storage hold the first information group and the second information group of the selected information set and then making the first storage delete the stored information set;
• determining whether or not each of the held first and second information groups satisfies an end condition and then adding processing completion information to the information group that satisfies the end condition;
• acquiring from the second storage, as the second and subsequent division processing target sets, the information groups which are held in the second storage and to which no processing completion information has been added;
• deleting, when the processing completion information has been added to each of the information groups held in the second storage, the information group of the subset that does not include at least N (N is a natural number) real examples from a plurality of the information groups held in the second storage;
• and
• acquiring all the information groups that are not deleted and remain in the second storage and outputting the acquired information groups.

A rule generation method comprising acquiring a training set including a plurality of examples, wherein the plurality of examples include at least one real example and at least one synthetic example, each of the real example and the synthetic example includes a feature value vector of which vector elements are one or more feature values corresponding to feature parameters different from each other, and a target value, and the feature value and the target value included in each of the real examples are measured values, while each of the synthetic examples is an example formed based on the real example,

• the rule generation method further comprising generating a rule group for dividing the training set into a plurality of clusters related to the target values using a rule base model so that a first constraint is satisfied, wherein each of rules in the rule group includes a condition obtained from a condition set including at least one simple condition related to the feature parameters and a predicted value related to the target values,
• wherein the first constraint includes a constraint that each cluster includes at least N (N is a natural number) real examples,
• wherein the generating the rule group comprises:
• generating a condition for each of the rules;
• and
• generating a predicted value of each of the rules, and
• wherein the generating the condition further comprises:
• selecting one of a plurality of explanatory variables respectively corresponding to a plurality of the feature parameters of examples included in a division processing target set and generating a plurality of thresholds using the selected explanatory variable, thereby generating a plurality of simple condition candidates for the selected explanatory variable;
• and
• sequentially selecting one simple condition candidate from among the plurality of generated simple condition candidates as a use simple condition candidate to be used for processing, forming a first subset and a second subset by dividing a plurality of the examples included in the division processing target set based on the use simple condition candidate, and outputting, for each of the simple condition candidates, an information set including a first information group in which the use simple condition candidate is associated with the first subset as a condition set and a second information group in which a negation of the use simple condition candidate is associated with the second subset as a condition set and then storing the information sets in a first storage,
• the first division processing target set is the training set,
• the generating the condition further comprises:
• selecting, from a plurality of the stored information sets, the information set in which each of the first and the second subsets satisfies the second constraint and a first statistical value calculated from the first and the second subsets is the smallest statistical value;
• making a second storage hold the first information group and the second information group of the selected information set and then making the first storage delete the stored information set;
• determining whether or not each of the held first and second information groups satisfies an end condition and then adding processing completion information to the information group that satisfies the end condition;
• and
• acquiring from the second storage, as the second and subsequent division processing target sets, the information groups which are held in the second storage and to which no processing completion information has been added, and
• the second constraint includes a constraint that each of the first and the second subsets includes at least N (N is a natural number) real examples.