Biclustering and Subspace Learning with Regularization for Financial Risk Analysis

摘要

Financial models draw on the need to turn critical (economical) information into better decision making models. When it comes to performance enhancement many advanced techniques have been used in bankruptcy detection with good results, yet rarely biclustering has been considered. In this paper, we propose a two-step approach based first on biclustering and second on subspace learning with constant regularization. The rationale behind biclustering is to discover patterns upholding instances and features that are highly correlated. Moreover, we placed great emphasis on building a weight affinity graph matrix and performing smooth subspace learning with regularization. In particular, the geometric topology of biclusters is preserved during learning. Experimental results demonstrate the success of the approach yielding excellent results in a real French data set of healthy and distressed companies.