The stability of different aggregation techniques in ensemble feature selection

Abstract To mitigate the curse of dimensionality in high-dimensional datasets, feature selection has become a crucial step most data mining applications. However, no method consistently delivers best performance across different domains. For this reason and order to improve stability process, ensemble frameworks have increasingly popular. While many examined construction techniques under various considerations, little work been done shed light on influence aggregation process selection. In contribution field, aims explore impact some selected strategies ensemble’s accuracy. Using twelve classification real datasets from domains, accuracy five were four standard filter methods. The experimental analysis revealed significant differences both behavior aggregations, especially between score-based rank-based strategies. Moreover, it was observed that simpler based Arithmetic Mean or L2-norm appear be efficient compelling cases. Given structure associated application domain, work’s findings can guide ensembles using suitable rules.