Low-rank Regularized Feature Representation Learning For Fashion Compatibility Prediction

The rapid spread of Internet and the booming online shopping platforms have made online shopping become a trend.With the development of artificial intelligence technology,fashion intelligent analysis can provide consumers with personalized services,such as predicting fashion trends,recommending matching items and scoring the compatibility of the outfit.In this context,with the continuous fashion pursuit and the daily dressing collocation problem of people,this paper focuses on the compatibility score prediction problem of fashion outfit.In many studies,there is a lack of analysis of outfit as a whole.Similarly,the capture of hidden feature patterns that affect fashion compatibility prediction,and the analysis of complex relationships between raw data are the entry points of problem.To solve the above challenges,two algorithms are proposed,which are Low-rank Hypergraph regularized Multiple Representation Learning(LHMRL)and Multi-graph Regularized Low-rank Projection Learning(MRLPL).LHMRL represents the original feature of fashion items from some different perspectives,while low-rank constraints remove redundant interference.The model is based on the Grassmann manifold theory to maximize the difference between multiple feature representations,combined with multiple hypergraph regularization constraints to model the complex relationship between fashion items and outfits.To enhance the representation ability of the model,two types of supervision information from labeled data are used,which are a multilabel matrix representing a relationship between items and outfits,and a supervised Lasso.MRLPL first decomposes the original feature matrix into a main feature part,a salient feature part and a sparse error part.The outfits are modeled by combining the main and the salient features with a binary associative matrix representing the relationship between items and outfits.To mine the complex structure that affects fashion compatibility prediction,model introduced three graph constraints to capture the latent structure of feature manifold,the sample manifold,and the outfit manifold.To improve the representation of model,we introduce supervised information to guide the learning of the model,using the Lasso method and considering the fashion compatibility prediction as a regression problem.The experiments on a publicly largescale dataset Polyvore and comparison with some existing classical algorithms and the latest algorithms demonstrate the superiority of our proposed model to the state-of-theart methods.