Common Mutual Information Selection Algorithm and Its Application on Combination Forecasting

Abstract

The subset selection of individual prediction methods is gradually becoming a hot topic. Among numerous forecasts, identifying the optimal subset approach has become a major focal point of research. To address this issue, the paper introduces a novel method based on information theory, which is called common mutual information (CMI) selection algorithm. This optimal subset selection method not only simultaneously considers the relationships of three factors, which include the candidate feature set, the selected feature set, and the actual time series, but also provides a more precise treatment of these relationships. Therefore, CMI algorithm employs the mutual information (MI) shared among the three factors as the criterion for selection and improves the accuracy of the redundancy or correlation measure for existing algorithms. Furthermore, it overcomes the deficiency of calculating MI between the candidate subset and the actual time series. Existing algorithms use the average MI values between individual elements within the subset and the actual sequence; this paper takes the selected subset as a multidimensional input for MI computation, thus reducing computational errors. Finally, the proposed algorithm is compared with two other approaches of the MI algorithm, the Max-Relevance and Min-Redundancy (mRMR) algorithm in both theoretical and empirical aspects. The experiments are illustrated to show the effectiveness and superiority of CMI algorithm.