Research And Implementation Of Recommendation Algorithm For Product Review Topics Based On Deep Learning

With the increasing advancement of big data and artificial intelligence technologies,the information data of various industries has shown exponential growth.Mass data can provide convenience,but it also brings prominent problems such as "information overload" and "information fragmentation".In the face of such problems,the current solutions include search engines and recommendation systems.An excellent recommendation system can not only provide users with more accurate and fast retrieval services,but also help businesses better locate potential target users.It can be seen that the in-depth study of recommendation algorithms is very important and has practical significance.In recent years,deep learning has been widely used in the field of personalized recommendation.Typical models include Deep Crossing,DIN,Deep FM,AFM,etc.Compared with the traditional recommendation model,the deep learning model has stronger expressive ability and can dig out more hidden features in the data.However,most recommendation models based on deep learning have some shortcomings.For example,(1)product review text is usually modeled as a single feature,and is not used to characterize the user's historical behavior;(2)changes in user interest in user reviews at different times It has important reference value,and the relationship between the two is not considered at present;(3)There are few researches on the fusion of image and text features.In response to the above problems,the main work of this article is as follows:First,a "Review Text Recommendation Model Based On Two-Layer Filtering Mechanism"(TFM)is proposed.This model introduces user comment timestamp and self-attention mechanism in the text feature processing part.For a given candidate item,by studying the correlation between the user's historical behavior and the item,it adaptively calculates the expression vector of the user's interest.Secondly,"Deep Learning Recommendation Model Based on Multimodal Feature Fusion"(MFM)is proposed.The model is divided into user side and item side.On the user side,it is proposed to use Transformer to pre-train the user comment text,which can alleviate the problem of feature information loss to a certain extent.On the article side,a feature aggregation module is proposed,which introduces the description text and image features of the article into the process of constructing the image of the article.Finally,using the Amazon product data set as the experimental data,through the design and construction of an online product recommendation system,online verification of the two proposed recommendation algorithms is carried out.The system comprehensively considers the user's cold start and item cold start issues,and combines big data components to complete online deployment and visual display effects.