| With the rapid development of the Internet,people are more and more inclined to buy the products they need online.Online shopping provides customers with a richer choice of goods and a more convenient shopping experience.In the context of the increasingly mature and standardized development of Internet e-commerce,competition among e-commerce platforms is becoming increasingly fierce.E-commerce platforms urgently need to improve service efficiency and quality,so as to improve customers’ shopping experience and increase their own competitive advantages.In the scenario of products recommendation and price comparison,e-commerce platforms need to establish a set of efficient and accurate products matching system.Products information often has multi-source modes.By integrating multi-source data,the accuracy of products matching model can be improved.In this paper,a products matching model based on image and text is constructed to achieve more accurate and efficient products matching.1.A set of products matching system is constructed,which is mainly divided into feature extraction and products matching.In feature extraction,ResNet18 model and Bert model are mainly used,and the parameters of the model are finetuned.2,In view of the products matching system faced with the problem of data sets is open,the researchers used in feature extraction model of fine-tuning the on the face recognition problem encountered similar problems by Arcface loss function,the extracted features significantly increased within the class of tightness,with Softmax contrast experiment of goods matching F1 score increased 0.075.3.A comprehensive distance calculation formula of text feature distance and image feature distance is proposed.While taking into account the sum of the distances of the two features,the more prominent distance features are matched first.A neighborhood hybrid method for hybrid feature matching is proposed,which us es the feature information of itself and the neighborhood node simultaneously to expand the ability of feature query and make the matched products feature vector more comprehensive.At the same time,the characteristic information of neighborhood node is used to change its own product feature vector,which makes its own feature vector more tend to the center of the class,and improves the matching accuracy.The experimental results show that the score of F1 is improved by 0.044 when using this method for feature matching compared with that without this method.Aiming at the practical problems of products matching,this paper optimized and innovated the feature extraction model selection,loss function,distance calculation and matching mode,which improved the model effect and realized more accurate and efficient products matching. |
