Research On The Prediction Of Sales By Online Product Graphical And Textual Information

As the main channel for consumers to understand products when they shop online,the information presented on product pages is decisive for consumers to make purchase decisions.The information on the product page includes both structured and unstructured data.The structured data is mainly the basic attributes of the product,such as price and category,which help consumers to get the basic information of the product,while the unstructured data is mainly the introductory images and text,which introduce the product more vividly and deeply and have stronger attraction to consumers.In order to investigate the impact of various types of information on product pages,especially graphical and textual information,on sales,the main work of this paper includes: Firstly,we use fully connected networks to integrate basic attribute features,and adopt the idea of transfer learning to extract the features of graphical and textual information on pages separately,and then compare the multimodal fusion methods based on decision layer and feature layer,the results show that the self-attention fusion in the feature layer fusion method achieves the best results on this task,but the efficiency is low.Secondly,we propose a fusion method based on step-by-step training based on the actual task and the model structure to make each feature extraction module get adequate training,and we find that following the sequence from simple to complex makes the step-by-step training method achieve the best performance,which is in line with people’s habit of obtaining information.The F1 score and accuracy score are improved by2.64% and 2.27% respectively compared with the self-attention fusion method,and the running time of the model is reduced by more than 20 minutes under the same training rounds,which significantly improves the training efficiency and proves the superiority of the method.Finally,the step-by-step training fusion method is used to add graphical and textual information to the basic attributes to compare their effects on sales prediction.The results show that both graphical and texual information on the pages of children’s books on Dangdang.com does have an impact on sales,and the effect of textual information is more obvious,and the impact of graphical and textual information on sales has an interactive enhancement effect.This paper enriches the research method of product page information,and finally proposes realistic suggestions for e-commerce platforms to optimize the page presentation to improve sales based on the findings.