Fuzzy factorization machine

Abstract

Rational and accurate classification cannot be achieved without considering both the historical information and domain knowledge. We propose fuzzy factorization machine (fuzzy FM) to integrate fuzzy set theory and factorization machine techniques for knowledge-enhanced classification. Each instance is assigned a membership through experts' estimations, and the instance's contribution to the objective function is weighted by its membership instead of the equal penalty in the standard FM. By adopting differentiated weighting strategies, we propose two variants of fuzzy FM: unilaterally weighted fuzzy FM (UFFM) and bilaterally weighted fuzzy FM (BFFM). In BFFM, each instance may not be fully assigned to one of two classes for better classification of imbalanced data, while in UFFM, each instance can only be assigned to one class. A set of membership generation approaches is summarized to quantify experts’ prior estimations. We introduce solving methods based on stochastic gradient descent for UFFM and BFFM. Experiments on real credit datasets demonstrate that the proposed fuzzy FM models can yield better rational classification than previous baselines (including the standard FM). The proposed fuzzy FM is a generic machine learning framework that can be applied to various rational classification tasks.