FSS-OBOP: Feature subset selection guided by a bucket order consensus ranking

Several authors have ton the importance of aggregating the results of different feature selection methods in order to improve the solutions obtained. To the best of our knowledge, the consensus rankings obtained in all of these proposals do not allow that some variables are tied. This paper studies the advantages of allowing ties in the consensus ranking obtained from aggregating several features selection methods. This implies that the consensus ranking is modeled as the problem of obtaining the Optimal Bucket Order instead of solving the Rank Aggregation Problem. In this paper we propose a filter-wrapper algorithm, that we will call FSS-OBOP, which uses a filter-based consensus ranking with ties to guide the posterior wrapper phase. By using a benchmark with 12 high-dimensional datasets, we show that allowing ties in the consensus rankings leads to subsets that, when used to induce a classifier, obtain at least the same, when not better, accuracy. Furthermore, and what is actually more significant, they reduce the number of wrapper evaluations extraordinarily.