A flexible framework for solving constrained ratio problems in machine learning

Abstract

The (constrained) optimization of a ratio of non-negative set functions is a problem appearing frequently in machine learning. As these problems are typically NP hard, the usual approach is to approximate them through convex or spectral relaxations. While these relaxations can be solved globally optimal, they are often too loose and thus produce suboptimal results. In this thesis we present a flexible framework for solving such constrained fractional set programs (CFSP). The main idea is to transform the combinatorial problem into an equivalent unconstrained continuous problem. We show that such a tight relaxation exists for every CFSP. It turns out that the tight relaxations can be related to a certain type of nonlinear eigenproblem. We present a method to solve nonlinear eigenproblems and thus optimize the corresponding ratios of in general non-differentiable differences of convex functions. While the global optimality cannot be guaranteed, we can prove the convergence to a solution of the associated nonlinear eigenproblem. Moreover, in practice the loose spectral relaxations are outperformed by a large margin. Going over to constrained fractional set programs and the corresponding nonlinear eigenproblems leads to a greater modelling flexibility, as we demonstrate for several applications in data analysis, namely the optimization of balanced graph cuts, constrained local clustering, community detection via densest subgraphs and sparse principal component analysis.