Table 3.1: Performance comparison of different approaches in terms of AUC

and bottoms, respectively. We noticed that the popular fashion items are all in the basic style,

such as plain T-shirts and jeans, which maybe due to the fact that they can go with many other

items. erefore, we can easily find the limitations of POP method. For example, most of the

popular bottoms are jeans, which maybe not suitable for professional tops and sport outfits.

erefore, it is not advisable to adopt recommendation strategy based on popularity. (2) ExIBR

and BPR-DAE both outperform the visual-based baseline IBR, which confirms the necessity

of considering the contextual modality in compatibility modeling. (3) BPR-DAE shows su-

periority over ExIBR. One possible explanation is that the highly sophisticated compatibility

space would be better characterized by the autoencoder neural networks rather than by the linear