Outlier Item Detection in Fashion Outfit

Abstract

In this paper, we introduce the outlier item detection task, which is related to the compatibility prediction. Although, with the ability of measuring the compatibility, we are able to identify items that do not match the overall style of a given outfit, the outlier item detection task has not been well studied before. Most existing methods on compatibility prediction focus on improving the recommendation accuracy by utilizing the underlying high order relationships among items and have achieved promising results. Since these methods are not designed to address the above problem, the performance can be relatively poor. In this paper, we introduce the outlier item detection task and propose an attention-based encoder to learn a permutation equivariant transformation for items. The encoder is independent of the size of the items. An MLP decoder is deployed to detect the outlier item. We conduct experiments on different fashion datasets and the empirical results show that our model achieves superior performance over the state-of-the-art methods.