Research On Sequential Recommendation Based On Self-attention Mechanism

Recommendation system is an effective tool to alleviate current information overload problem,and has become the core technology of e-commerce,short video,social platform and other Internet companies to accurately push products and services to users.Sequential recommendation algorithm is a subfield of recommendation algorithm,which aims to model the dynamic changes of user preferences according to the user’s historical behavior records,so as to predict the user’s next interested items and provide users with a recommendation list.With the rapid development of deep learning technology,more and more deep learning models are applied in the field of recommendation system.In recent years,the self-attention mechanism has achieved good performance in many fields.Researchers in the field of sequential recommendation applied it to the field of sequential recommendation,and the model also showed excellent results.However,the existing researches still have the following shortcomings: In the sequential recommendation model based on self-attention mechanism,the embedding representation of the item and the location information of the item are directly added as the input of the model,which will introduce noise when the model extracts features,and then affect the recommendation accuracy.The data sparsity problem makes it difficult to capture the relationship between some items.Aiming at the above problems,the main research contents of this paper are as follows:(1)Aiming at the problem that the direct addition method used by the sequential recommendation model based on self-attention mechanism when integrating the location information of the item will introduce noise in the feature extraction of the model,a nonintrusive self-attention mechanism is proposed and a sequential recommendation model based on non-intrusive self-attention mechanism(NISASRec)is designed.The nonintrusive self-attention mechanism extracts the features of two parts of input information through two self-attention mechanisms,avoiding the introduction of noise when calculating the attention score,and then improving the recommendation accuracy.(2)Aiming at the problem that it is difficult to capture the relationship between some items due to data sparsity,this paper introduces the category information of the item to enhance the ability of the model to capture the relationship between items.In real life,the category information of the item is often represented in the form of multi-label discrete,so a processing module of multi-label discrete category information is designed to process the category information of the item.Based on the NISASRec model,the category-aware non-intrusive embedding in sequential recommendation(CA-NISASRec)is designed.Experiments were conducted on three real datasets and compared with other baseline models.The evaluation metrics NDCG of NISASRec model were improved by 3.9% on average and 7.6% at the highest.The evaluation metrics NDCG of CA-NISASRec model were improved by 4.5% on average and 13.3% at the highest.Figure [18] Table [8] Reference [61]...