A Novel Information-based Multi-view Representation Learning

Multi-view representation learning methods achieve great performance in various domains via fusing complementary and consistent information of views, which have gained great attention. However, there still exist two issues in current methods. They typically assume strict semantic consistency across views to learn representations of multi-view data. Moreover, they lack a unified theoretical framework in mining patterns of multi-view data, making it difficult to gain a deeper understanding of multi-view representation learning. To this end, a new information-based multi-view representation learning within the encoding-decoding architecture is proposed to aggregate complementary and consistent information for mining patterns (IMVRL). It consists of three mutual information objectives for multi-view representation learning, i.e., concentration learning, consistent learning, and comprehensiveness learning. Three objectives work seamlessly in aggregating complementary and consistent information to mine patterns of multi-view data. Finally, a series of extensive tests across three datasets underscore the advantages and efficacy of IMVRL.