Aggregated local models via subspace clustering

Abstract

Decision models take a valuable step towards harnessing the problem of efficient risk assessment in social and technological environments. In particular, in bankruptcy prediction models it becomes difficult to know exactly what happens when so many financial and external variables are at stake. To partly tackle this problem, a new approach encompassing aggregated local models obtained via subspace clustering and intelligent decision technologies is proposed in this paper. The approach first takes co-clusters of firms and financial ratios found by a biclustering algorithm; second the weight affinity graph matrix embedding data points is built for learning the subspace clustering model; finally, a large margin binary classifier over the regularized model is used to make predictions on financial real data. We empirically show that our model (by combining biclustering with subspace learning) significantly outperforms the competing approach without biclustering and the alternative without subspace learning in terms of prediction accuracy without a significant increase in the computational cost. Furthermore, we propose a consensus of found local models which is able through a simple aggregate rule to improve results even further.