Research On Multimodal Learning To Rank Based On Deep Semantic Features

With the rapid development of computer information technology,shopping online has gradually entered our daily life and affected people's shopping habits.The rapid development of online shopping not only brings huge benefits to electricity supplier but also takes much convenience for consumers.For example,we can make shopping online anytime and anywhere,conveniently and quickly.However,because of the increasing variety and similar types of products,users face a large number of search results as well as faced with product labels that are mixed in high rankings or low correlations words,which results in low search accuracy,rough product categories,search results pages display confusion and other issues.Faced with the above problems,this paper studies the common information retrieval model from the traditional information retrieval field and compares the similarities and differences among Pointwise,Pairwise and Listwise.Based on the research of the three learning ranking models,in order to improve the accuracy of the model for product search,this paper creatively introduces the meaning of "product image semantics" as well as integrates "feature engineering" and "learning".Secondly,the paper study ListNet method in the Listwise model,and the method that transforms the fractional sequence into the probability distribution in the calculation process is further analyzed.Through optimization,we determine that the ListNet method is the most direct and effective method to solve the sort learning process in the product search process.After determining the target method of ListNet,combining the actual characteristics of the products search and the ListNet method ignoring the self-defect of the document location information in the entire sequence,this paper innovatively introduces "location-weighted loss factors " and SHF-SDCG loss function fusion framework through the experimental optimization.This paper intergrates position weight loss factor and ListNet loss function,and then uses neural network model and gradient descent method to optimize the model after the loss function fusion.Finally,we verify the validity and accuracy of the model by a series of experiments.In the final experiment,after "many rounds" optimization the new ListNet model was trained and tested by using dataset one(no image semantic feature)and dataset two(integrated image semantic feature),then we respectively.use NDCG@K and P@K indexes to evaluate the experiment results.In the last,the results show that the new ListNet model which introduces the semantic features of product images has more advantages and is of great guiding significance for the development of e-commerce.