Heterogeneous ensemble selection for evolving data streams

Ensemble learning has been widely applied to both batch data classification and streaming classification. For the latter setting, most existing ensemble systems are homogenous, which means they generated from only one type of model. In contrast, by combining several types different models, a heterogeneous system can achieve greater diversity among its members, helps improve performance. Although have achieved many successes in it is not trivial extend them directly stream setting. this study, we propose novel HEterogeneous Selection (HEES) method, dynamically selects an appropriate subset base classifiers predict under We inspired observation that well-chosen good may outperform whole system. Here, define candidate as expresses high predictive performance but also confidence prediction. Our selection process thus divided into two sub-processes: accurate-candidate confident-candidate selection. accurate context classifier with accuracy over current concept, while confident score higher than certain threshold. first sub-process, employ prequential estimate at specific time, new measure quantify provide method learn threshold incrementally. The final formed taking intersection sets classifiers. Experiments on wide range streams show proposed achieves competitive lower running time comparison state-of-the-art online methods.