A Feature Incremental Learning Method Based on Evidential Reasoning Rule
Li Tu, Ruirui Zhao, Jianbin Sun, Jiang Jiang · 2022
Traditional data mining methods are no longer applicable with the rapid growth of data scale and complexity, while incremental learning can update existing model incrementally according to continually arrived data streams. But there are few studies focusing on feature increment, that is, the dynamic arrival of new features. Evidential Reasoning Rule (ER Rule) classifier possesses comparative classification performance, and has the advantages of few information loss and uncertainty transferring. Therefore, a feature incremental learning method based on ER Rule is proposed. Firstly, the feature incremental problem studied in this paper is analyzed. Then, the feature incremental learning method is established through three parts, that is, data imbalance analysis, base classifier construction and base classifiers ensemble. Sepcially, the ensemble weight of each base classifier is calculated based on Mean Square Error. Finally, the performance of the proposed method is verified through various experiments.