Scaling Up Feature Selection: A Distributed Filter Approach

Traditionally, feature selection has been required as a preliminary step for many pattern recognition problems. In recent years, distributed learning has been the focus of much attention, due to the proliferation of big databases, in some cases distributed across different nodes. However, most of the existing feature selection algorithms were designed for working in a centralized manner, i.e. using the whole dataset at once. In this research, a new approach for using filter methods in a distributed manner is presented. The approach splits the data horizontally, i.e., by samples. A filter is applied at each partition performing several rounds to obtain a stable set of features. Later, a merging procedure is performed in order to combine the results into a single subset of relevant features. Five of the most well-known filters were used to test the approach. The experimental results on six representative datasets show that the execution time is shortened whereas the performance is maintained or even improved compared to the standard algorithms applied to the non-partitioned datasets.