| In recent years,researchers have proposed a series of recommendation models and intelligent recommendation systems in order to cope with the explosive growth of information volume in the Internet every day and to mine users’ interests in the massive data.With the development and introduction of deep learning,researchers have further researched recommendation models,among which mining users’ interest preferences from user behavior sequences for recommendation is the focus of attention today.The user behavior sequence is embodied in the user’s click record,rating record,purchase record and other sequence information of historical browsing items,from which the dynamic changes of user interests can be mined,and the content that users like can be recommended more accurately.Most of the existing recommendation models based on deep learning usually pool the recent user behavior sequences or use a recurrent neural network to extract the abstract representation of the sequence as the user’s interest preference,which is not sufficient for mining longer behavior sequences and difficult to capture the inherent relationship between user behavior sequences and user preferences.In response to the above problems,the main work as follows.（1）Based on the recurrent neural network RNN and attention mechanism,a recommendation system recall model LSIN（Long and Short Interest Network）is proposed that integrates long and short-term user behavior.The model divides the user behavior sequence into long-term and short-term sequences.It uses a recurrent neural network to encode,and then combines the user attribute features,and uses the attention mechanism to extract the user interest features in the short-term and long-term user behavior sequences,so that the model can be recalled in the recommendation system stage to further mine the user’s interest preferences.（2）Based on the Transformer model,a recommendation system ranking model UBIN（User Behavioral Interactive Network,UBIN）is proposed.It uses the Transformer model to encode the user’s behavior sequence,and then uses the attention mechanism to extract the user’s interest preference by combining the features of the user and the recalled candidate items,so that the model can give a more accurate recommendation list in the ranking stage of the recommender system.（3）Comparing experiments with other baseline recall and ranking models on Movie Lens-1M and Amazon-Food datasets,in which the recall rate and hit rate of the LSIN model on the Amazon-Food dataset are improved compared to the SDM model by 1% and 1.2%.The AUC values of the UBIN model on the two datasets are improved by 2% and 3% respectively relativing to DIN.In addition,the effect of sequence long queues on the LSIN and UBIN models and the influence of the main model parameters on the experimental results are also explored.（4）Based on models proposed in this paper,a papers recommendation system is designed and developed.The system collects and processes the relevant information of papers from the Papers With Code website,firstly uses the recall model LSIN that integrates the user’s long-term and short-term behaviors to recall the top dozens of papers that users are most interested in from the dataset,and then uses the ranking model UBIN sorts the recall result list according to the click-through rate score,and finally returns the paper list to the user.In addition,it also develops functions such as recommendation for you,homepage recommendation,and similar items recommendation to provide researchers with better paper recommendation services. |
