Applying Class-to-Class Siamese Networks to Explain Classifications with Supportive and Contrastive Cases

Abstract

Case-based classification is normally based on similarity between a query and class members in the case base. This paper proposes a difference-based approach, class-to-class siamese network (C2C-SN) classification, in which classification is based on learning patterns of both similarity and difference between classes. A C2C-SN learns patterns from one class to another class. The network can then be used, given two cases, to determine whether their similarity and difference conform to the learned patterns. If they do, it provides evidence for their belonging to the corresponding classes. We demonstrate the use of C2C-SNs for classification, explanation, and prototypical case finding. We demonstrate that C2C-SN classification can achieve good accuracy for case pairs, with the benefit of one-shot learning inherited from siamese networks.