关键词: 增量学习, 集成学习, 局部敏感哈希, 异构分类器集成, 动态权重

Abstract:

Introducing the idea of ensemble learning into incremental learning can improve the learning effect. In recent years, most of the research on ensemble incremental learning combines multiple homogeneous classifiers with weighted voting method, which does not solve the problem of stability-plasticity in incremental learning very well. An incremental learning algorithm based on heterogeneous classifier ensemble is proposed. In the stage of training, to make the ensemble model more stable, many base classifiers are trained with new data and then append into heterogeneous ensemble model. Meanwhile, Locality-Sensitive Hashing is used to save the data sketch for the nearest neighbor search of test sample. In order to adapt to the changing data, the newly acquired data will be used to update the voting weight of the base classifier in the ensemble model. In the prediction stage, for the class label prediction of the test sample, the data similar to the test sample is found from the Local-Sensitive Hashing table, this data is used as a bridge to calculate the dynamic weight of the base classifier for the test sample. It determines the class label of the test sample by combined the voting weight and dynamic weight of many base classifiers. Through comparative experiments, it is proved that the proposed algorithm has well stability and generalization ability.

Key words: incremental learning, ensemble learning, Locality-Sensitive Hashing（LSH）, heterogeneous classifiers ensemble, dynamic weight