Abstract
With the rapid development of the online shopping, the pursuit of outfit compatibility has become a basic requirement for an increasing number of customers. The existing work on outfit compatibility prediction largely focuses on modeling pairwise item compatibility without considering modeling the whole outfit directly. To address the problem, in this paper, we propose a novel hypergraph-based compatibility modeling scheme named OCPHN, which is able to better model complex relationships among outfits. In OCPHN, we represent the outfit as a hypergraph, where each hypernode represents a category and each hyperedge represents the interactions between multiple categories (i.e., they appear in the same outfit). To better predict outfit compatibility, the hypergraph is transformed into a simple graph, and the message propagation mechanism in the graph convolution network is used to aggregate the neighbours’ information on the node and update the node representations. Furthermore, with learned node representations, an attention mechanism is introduced to compute the outfit compatibility score. Using a benchmark dataset, the experimental results show that the proposed method is an improvement over the strangest baselines in terms of accuracy by about 3% and 1% in the fill-in-the-blank and compatibility prediction tasks, respectively.
Funder
Tianjin University
Tianjin Municipal Science and Technology Bureau
Subject
General Mathematics,Engineering (miscellaneous),Computer Science (miscellaneous)
Reference48 articles.
1. Visually-Aware Fashion Recommendation and Design with Generative Image Models;Kang;Proceedings of the 2017 IEEE International Conference on Data Mining (ICDM),2017
2. Interpretable Partitioned Embedding for Customized Multi-item Fashion Outfit Composition;Feng;Proceedings of the 2018 ACM on International Conference on Multimedia Retrieval,2018
3. Creating Capsule Wardrobes from Fashion Images;Hsiao;Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR),2018
4. Multi-Modal and Multi-Domain Embedding Learning for Fashion Retrieval and Analysis
5. What Aspect Do You Like: Multi-scale Time-aware User Interest Modeling for Micro-video Recommendation;Jiang;Proceedings of the 28th ACM International Conference on Multimedia,2020
Cited by
1 articles.
订阅此论文施引文献
订阅此论文施引文献,注册后可以免费订阅5篇论文的施引文献,订阅后可以查看论文全部施引文献