Deep Fashion Recommendation System with Style Feature Decomposition
Yong-Goo Shin, Yoon-Jae Yeo, Min-Cheol Sagong, Seo-Won Ji, Sung-Jea Ko · 2019
Recently, convolutional neural network (CNN) -based fashion recommendation techniques, which automatically recommend the matching clothes to the consumer, have been widely researched. In general, the feature vector of a fashion item, i.e. clothes vector, obtained by CNN conveys two types of information: style and category, where the style indicates the distinctive characteristic of the clothes and the category represents the common properties of the clothes in the same class. Due to the mixed information of style and category, however, the clothes vector often recommends the unmatching clothes. To solve this problem, we propose a style feature extraction (SFE) layer, which effectively decomposes the clothes vector into style and category. Based on the characteristics that the category information has small variations in the same class while being distinguished from other classes, we extract and remove the category information from the clothes vector to obtain more accurate style information. Experimental results show that the proposed method achieves state-of-the-art results in terms of link prediction, which is a performance measure of a stylish match. In addition, as a simple CNN layer, it is expected that the proposed SFE layer is compatible with all popular CNN architectures.