Correlation-Oriented Graph Learning for OCM

In this chapter, we study how to automatically evaluate the compatibility of a given outfit that involves a variable number of items with the graph learning technique. Existing graph learning-based methods focus on exploring the visual modality of fashion items, and seldom investigate an item’s textual aspect, i.e., the textual description. In fact, textual descriptions of fashion items usually contain key features, which benefit item representation learning. Notably, although some studies have attempted to incorporate the textual modality, they simply adopt early/late fusion or consistency regularization to boost performance. Nevertheless, the correlations among multimodalities are complex and sophisticated and are not yet clearly separated and explicitly modeled.