An Evolutionary Rule Mining Method for Continuous Value Prediction from Incomplete Database and Its Application Utilizing Artificial Missing Values

A rule mining method for continuous value prediction has been proposed to handle incomplete databases using a graph structure-based evolutionary computation technique. The method extracts the associative local distribution rule, the consequent part of which has a narrow distribution of continuous variables. A set of associative local distribution rules was applied for continuous value prediction. Instances including missing values were predicted using the predictor. A method for constructing a probability distribution of predicted values for each focusing instance was considered based on extracted rule sets. The proposed method offers some flexibility by allowing users to define the conditions of prediction rules. The method can quit rule extraction when a sufficient number of rules are extracted for building a predictor. Therefore, it is suitable for prediction when large datasets are involved. In addition, we have proposed an application of artificial missing values to improve the effectiveness of the developed rule-based prediction system. Artificial missing values are applied to avoid the sharp boundary problem encountered when discretizing continuous variables. Attribute values near the boundary in discretization are treated as missing values. The performance of the artificial missing value-based prediction method was evaluated, and the results showed that the proposed method was effective for prediction.