📝 SummarXiv

Abstract

This paper addresses the challenge of determining optimal cut-offs for a set of n items with m scores to maximize distinguishability. The term distinguishability is defined as the fraction of item pairs assigned to different buckets, where buckets are determined by the number of cut-offs an items scores exceed. A brute-force approach to this problem is computationally intractable, with complexity growing exponentially with the number of scores. On the other hand, attempts to solve the problem in the continuous domain lead to local minima, making it unreliable. To overcome these challenges, the problem is formulated as a Integer Quadratic Program (IQP). Since IQPs become computationally difficult with even moderate size problems, a surrogate Integer Linear Program (ILP) is introduced, which can be solved more efficiently for larger instances. In addition to these exact methods, a simple heuristic is proposed that offers a balance between solution quality and computational efficiency. This heuristic iteratively adjusts cutoffs for each score, considering a finite-set of meaningful cut-off points. Computational results provide an empirically evidence for the effectiveness of our proposed methods.