Evaluating Multiple-Concept Biomedical Hypotheses Based on Deep Sets

Juncheng Ding, Wei Jin

Abstract: Given hypotheses that connect two “irrelevant” concepts of interest via one or multiple concepts, Hypothesis Generation (HG) attempts to judge their meaningfulness and compare or rank them accordingly. HG can accelerate scientific research and is becoming increasingly important. The basic idea of prior studies is to conduct a two or higher-order search between the two input concepts for the one-concept and multiple-concept hypotheses to evaluate them. However, these approaches inevitably encounter exponential-growing searching space when addressing multiple-concept hypotheses, making it impractical to tackle such hypotheses. We propose HG Set Net (HSN) that forms a hypothesis with any number of connecting concepts as a set and learns to evaluate the set to address this problem. HSN can evaluate any hypotheses with the same complexity and avoid higher-order search, making it computationally possible to evaluate hypotheses with numerous concepts. Besides, we present a double-margin loss to train HSN to resolve the lack of labeled hypotheses. Experiments show that HSN can not only address hypotheses with efficiency but also outperform previous approaches. The double-margin loss also reveals to boost HSN’s performance.