Research On Personalized Outfit Recommendation With Pseudo User Behavior And Feature Factorization

With the rapid development of the online fashion industry,Internet communities and e-commerce platforms provide a large amount of clothing for users to choose from.The massive amount of information makes it difficult for users to quickly find items they are interested in.Research on personalized recommendation algorithms has surged to help users efficiently filter target items according to their interests.However,existing recommendation algorithms only focus on single-item recommendations based on users’ interests in items.In actual e-commerce scenarios,users often buy matching items together when purchasing a single item.Therefore,outfit recommendation has become a new research hotspot.Unlike single-item recommendation,outfit recommendation needs to consider the combination relationship between the items in the outfit and the correlation between the user and the outfit.However,current research methods have certain limitations.Firstly,the user expression based on ID embedding and single-visual item representation is not sufficient.Secondly,the feature expression of the unified space cannot model the multi-angle interests of users and the multi-attribute characteristics of items.To address these issues,this paper proposes two types of research:(1)To address the first issue,this paper constructs a multi-modal outfit recommendation model based on pseudo-user behavior sequences,which strengthens the representation of users and outfits in three aspects.Firstly,pseudo-user behavior sequences are constructed based on the minimum variance sampling method to mine user interests from historical behaviors to strengthen user representation.Secondly,the stacking structure of the attention mechanism is used to perform high-order combination extraction features on the outfit to strengthen the outfit representation.Finally,the text vision multimodal feature fusion method is used to complement the modal information to further strengthen the representations.(2)To address the second issue,this paper builds an outfit recommendation model based on feature factorization,which extracts the multi-angle representation of users and outfits from two aspects.On one hand,the feature factorization of users and outfits is carried out to represent the multi-angle interests of users and the multiattribute features of outfits;On the other hand,a factor matching module is built to learn the corresponding combination relationship between the multi-angle interests of users and the multi-attribute features of outfits from a fine-grained perspective.Extensive experiments on the Polyvore-630 and Polyvore-519 datasets show that the proposed recommendation model has effectively improved the performance.Additionally,the method proposed brings inspiration to the follow-up research of feature enhancement and feature factorization.