Research On Fine-grained Fashion Compatibility Prediction Model Based On Attention Mechanism

With the increase of social occasions,people are pursuing beautiful,fashionable and appropriate outfits.Building intelligent clothing matching algorithms is important to meet user needs and the development of the fashion industry.Clothing fine-grained compatibility modeling is a challenging task in terms of how to fuse multi-modal information of fashion items to extract their compatible interaction features,and use them to determine whether there is a compatibility relationship between clothing items,as while as apply them to personalized recommendation tasks.In this paper,we address the above issues and our work includes the following two aspects:(1)Personalized compatibility embedding network based on multi-modal attention(PCENet)is proposed,which consists of two components: a compatibility embedding component based on multi-modal attention branches and a personal preference component based on matrix factorization.First,the visual and textual compatibility interaction relations of clothing are learned by designing multi-modal attention branches.Secondly,personal preferences of users are obtained based on matrix factorization and content features.Finally,these two parts are jointly trained based on Bayesian Personalized Ranking framework(BPR).This approach solves the problem of the utilization of multi-modal features in clothing compatibility prediction algorithms.Extensive experiment results on the existing clothing dataset IQON3000 show that the effectiveness of multi-modal attention branches for the compatibility prediction task,and PCE-Net outperforms most baseline models on the personalized clothing recommendation task.(2)Learning similarity condition subspaces embedding network based on mutual attention(MASCE-Net)is proposed.Existing single embedding models ignore the fine-grained attribute interactions between clothing data.This paper focuses on this problem to build a subspace embedding network by designing a mutual attention block,and obtain the relative importance of clothing compatibility interaction features in the fine-grained attribute subspace,so as to obtain the fine-grained attribute features between clothing images and evaluate whether there is a compatibility relationship.This approach compensates for the shortcomings of the clothing compatibility algorithm in learning fine-grained compatibility interaction features.Extensive experiments on two datasets,IQON3000 and Zappos50 K,validate the effectiveness of the method for the tasks of triplet clothing compatibility prediction and fine-grained similarity evaluation.