Research On Clothing Classification And Collocation Algorithm Based On Heterogeneous Network

In recent years,with the rapid development of e-commerce,buying clothing online has become a popular trend.Many researchers have carried out a series of studies on clothing classification,attribute prediction,clothing retrieval,clothing collocation and other topics in the fashion field,which has created important commercial value for the industry.Among these topics,clothing classification and collocation are hot research problems in the fashion field,whose results greatly affect users' shopping experience.In the real-life scene,there are certain relationships between various types of clothes,including similar and matching relationships between clothes.In the existing studies,features of the clothes themselves,including image features and text features,are often used to solve the problems of clothing classification and collocation,but relationships between clothes are ignored.Aiming at the above problem,this paper constructs a fashion heterogeneous network with similar and matching relationships between clothes to solve the problems of clothing classification and collocation from a new perspective.The paper uses the clothing dataset from an e-commerce website to construct a fashion heterogeneous network.In the network,nodes represent clothes and edges represent similar and matching relationships.After that,the network representation learning methods are used to convert the fashion heterogeneous network into vectorized representations to realize downstream task of clothing classification.The paper proposes an improved Relational Graph Convolutional Network(R-GCN)clothing classification model and compares it with other network representation learning models.The improved R-GCN classification model proves to be effective in the task of clothing classification.In addition,by using different numbers of R-GCN layers and R-GCN hidden layer units for comparative experiments,the paper determines the structure of the encoder in the classification model based on R-GCN.The paper realizes matching structural feature representing based on General Attributed Multiplex Heterogeneous Network Embedding(GATNE)and matching textual feature representing based on siamese network.The paper further proposes a clothing collocation model based on the fusion of matching structural feature and textual feature and evalutes it by comparing with the clothing collocation model based on GATNE and the clothing text collocation model based on siamese network.The experimental results show that the clothing collocation model based on the fusion of matching structural feature and textual feature is better than both the clothing collocation model based on GATNE and the clothing text collocation model based on siamese network.The matching accuracy of the proposed model is so high that it is suitable for actual clothing collocation scene.