Unsupervised Disentangled Graph Learning for OCM

In Chapter 2 and Chapter 3, we proposed two methods for outfit compatibility modeling. They still suffer from two key limitations. (1) They evaluate the outfit compatibility based on the single latent compatibility space. The outfit compatibility is essentially affected by multiple complementary hidden factors, such as the color, style, shape, and material. Therefore, we argue that previous methods can only achieve the suboptimal solution, as it entangles all the factors in a single latent space. (2) They focus on learning the representation of each composing item and based on that, calculate the outfit compatibility. We argue that this method still fails to authentically treat the outfit as a whole, namely, it overlooks the global outfit representation learning. Therefore, in this chapter, we aim to estimate the compatibility of the outfit by considering the multiple hidden spaces and the global outfit graph representation learning.