Clothing Matching Recommendation Based On Graph Neural Network

Under the background of "fast fashion",clothing matching technology based on deep learning has become a hot topic in the research fields of personalized fashion design,fashion retrieval and so on.Because the key features such as color depth and texture pattern of different clothes are quite different,and different user groups have different requirements for clothing styles,the existing clothing matching technology based on deep learning has some problems,such as incomplete representation of clothing features,low correlation of multi-modal feature fusion,unreasonable matching degree evaluation of outfits,poor matching degree of complementary items,and low recommendation efficiency.This paper focuses on the research of clothing matching recommendation algorithm,and carries out the following work:1.A learning method of personalized clothing match modeling based on multi-modal fusion is realized.In this method,in addition to taking the multi-modal information of traditional items as the recommendation condition,user factors are creatively introduced into the modeling method to realize personalized garment recommendation for different user groups.Through the experiment on IQON3000 data set,the accuracy of the model is improved by 1.25% compared with the traditional method.2.A multi-modal clothing match modeling learning method based on graph neural network（MMOR）is designed.Firstly,aiming at the problem that the representation of clothing features in this method is not complete,in the multi-modal clothing matching modeling MMOR network based on graph neural network proposed in this paper,FPN and Text CNN networks are used to extract the image and text features of items and obtain the potential representation of multi-modal features of items,Thus,it lays a good foundation for improving the ability to capture the subtle features of items during model training.3.Use the potential representation of items and category co-occurrence matrix to build a outfit graph to solve the problem of low correlation of feature fusion.The category co-occurrence matrix adaptively distributes the weight of the edges in the outfit graph,and uses the message passing mechanism to introduce the aggregation module between intra-modal and inter-modal to aggregate the fine-grained information in the item inter-modal and the coarse-grained information in the item intra-modal,so as to enhance the correlation of multi-modal feature fusion of items in the set.4.In order to improve the rationality of outfit matching degree evaluation and the matching degree of complementary items,this paper uses the multi head attention mechanism to distinguish the importance of clothing items in the outfit,according to the items represented by multi-modal feature fusion,and measures the overall matching degree of complementary items in the outfit,which better solves the problem of unreasonable outfit matching degree evaluation.The experimental results show that compared with traditional method,the AUC and ACC indexes of MMOR network embedded with multi head attention mechanism,intra-modal and inter-modal aggregation modules on Polyvore-D dataset are improved by7.14% and 13.40% respectively,and the performance on Polyvore-ND dataset is also greatly improved.