Research On Long And Short-term Neural Network Recommendation Model Based On Self-attention Mechanism

In recent years,with the advent of the era of big data and the rapid development of 5G technology,online news platforms such as Net Ease News,Toutiao News and Tencent News have become important media for people to obtain information.While enjoying the convenience of information,information overload has gradually become a huge barrier for people to efficiently obtain information.The personalized news recommendation system can help users browse news that meet their personal preferences through a large amount of news,effectively alleviate the problem of information overload and it can also help news platforms improve user experience and user stickiness.Existing news recommendation systems based on deep neural networks usually process the user's past browsing history in a unified manner,ignoring the difference between the user's longterm preferences and short-term preferences,which will directly affect the accuracy of user behavior feature extraction.At the same time,how to accurately capture the phenomenon of user interest migration in the short term is also an urgent problem in the news recommendation system.In response to the above problems,the thesis studies a long and short-term memory model based on Gated Recurrent Unit(GRU)and proposes an improved long-and short-term memory model based on the self-attention mechanism.This paper uses Latent Factor Model(LFM)to extract user's longterm preferences and uses GRU to obtain short-term preferences from user's browsing records.Aiming at the problem of user interest transfer in the short term,this paper adopts a self-attention mechanism based on time interval to characterize the degree of user interest transfer.The paper further studied the new model combining the above two ideas and verified through experiments that the long and short-term memory model based on the self-attention mechanism proposed in this paper improves the recommendation accuracy compared with the traditional long and short-term memory model.Specifically,the main research of this article is carried out from the following three aspects:(1)Compared with the method of directly initializing the hidden state of the GRU randomly in the general model,this paper chooses to initialize the GRU using the user's long-term preferences.When GRU is used to analyze the user's browsing records to obtain the user's short-term preferences,this approach can not only strengthen the influence of long-term preferences,but also optimize the training effect of the model.The model proposed in this paper uses Word2 Vec technology to constitute news features by analyzing news headline information,the first-level category information and the second-level category information which belong to the news.At the same time,a self-attention mechanism based on news attributes is added to adjust the weights of different news attributes and optimize the accuracy of news feature extraction.(2)The long and short-term memory model based on self-attention mechanism proposed in this paper obtains user's long-term preference characteristics and short-term preference characteristics.They are cascaded to facilitate predictions of user's click-through rate on candidate news.Two preferences are well distinguished.Besides,this paper proposes a self-attention mechanism based on time interval which alleviates the phenomenon of user interest transfer by considering the time interval information between two news.The attention mechanism can adjust the attention weight between two news that the user has already viewed,ensuring that the more distant the two news,the smaller the degree of correlation between them.(3)Based on two sets of real public data sets,the improved model proposed in this article is compared with four commonly used models in the news recommendation field.The experimental results show that the long and short-term memory model proposed in this paper based on selfattention mechanism has good recommendation effects on different data sets.