A Study on the Personal Fashion Preference in Social Media using Meta-path based Heterogeneous Graph Modeling

Abstract

As fashion is subjective, user preferences need to be taken into account when recommending fashion items. In this study, we introduce Graph-based Look Attribute preference Modeling (GLAM), a heterogeneous graph deep learning model that utilizes outfit postings on social media to understand user preferences. GLAM consists of (C1) data collection, (C2) fashion object detection, (C3) fashion item clustering, and (C4) graph model training. Data was obtained from the representative online fashion community Lookbook.nu, including 2,497 users’ information, 456,329 outfit images, and 328 extracted attributes. A heterogeneous graph was constructed with the nodes representing users, item images, and fashion attributes. Through metapath-based graph learning, the multi-modal node features were updated reflecting relations between node types. We evaluate GLAM’s performance in recommending fashion items based on user preferences. GLAM demonstrates comparable performance to existing graph-based models, with the potential for providing explanatory captions in recommendations based on fashion attributes.