Abstract:

The relevance between the candidate feature and the class label, the interaction information among the candidate feature, the selected feature and the class label, and the redundancy between the candidate feature and the selected feature are important factors that should both be considered by feature selection algorithms. Some feature selection algorithms based on mutual information and three-dimensional mutual information do not consider the relevance, the interaction information and the redundancy at the same time, which affects their performance. Therefore, a feature selection algorithm based on dynamic weights using redundancy is proposed. The algorithm uses three-way interaction information and symmetrical uncertainty as criteria and adopts a method for dynamically updating the weights of candidate features. The objective function can emphatically consider the redundancy between the candidate features and the selected feature on the basis of the fact that the relevance and interaction information are considered. Comparative experiments with typical feature selection algorithms based on mutual information are conducted on ten datasets by using three classifiers. The experimental results show that the proposed algorithm has a better feature selection performance.

Key words: feature selection, redundancy, three-way interaction information, symmetrical uncertainty, classification