Research Frontiers

Thus far, in this book, we studied the task of outfit compatibility modeling, where each outfit involves a variable number of items. In particular, we first identified the prominent research challenges we faced to solve this task, including the multiple correlated modalities, complicated hidden factors, nonunified semantic attributes, and users’ personal preferences. To address these challenges, we proposed a series of graph learning theories. In particular, we first presented a correlation-oriented graph learning method for outfit compatibility modeling, which explicitly models the consistent and complementary relations between different modalities (i.e., the visual and textual modalities). Considering that this scheme overlooks the category modality and the intermodal compatibility modeling, we next introduced a modality-oriented graph learning method for outfit compatibility modeling. Beyond these two methods that focus on the coarse-grained compatibility modeling, we then devised an unsupervised disentangled graph learning method to uncover the hidden factors affecting the overall compatibility and fulfill the fine-grained compatibility modeling. Moreover, to fully utilize item-attribute labels, we further developed a partially supervised disentangled graph learning method. Finally, to incorporate the user’s personal tastes, we proposed a metapath-guided heterogeneous graph learning scheme for personalized outfit compatibility modeling.