Research On Sequence Recommendation Based On Transformer

In the era of big data,information is exploding,fast and accurate access to effective information has become a new challenge,to solve this challenge,the recommendation system have been created.As an effective information filtering tool,it can help people select the information from the Internet that meets their own intention.Research related to traditional recommendation models to sequential recommendation models has been widely paid attention by researchers.Traditional recommendation methods model the user's historical behavior statically,considering only the user's general preferences.Sequential recommendation models attempt to dynamically model the user-item interaction records,and it can achieve accurate and personalized recommendation by constructing a sequential pattern.Existing recommendation systems take the order of user-item interaction as a key feature to obtain the next behavior of the user.In previous studies,Markov chains and recurrent neural networks are the two main applied methods,and both of them have certain drawbacks in dealing with datasets with different sparsity.In addition,sequential recommendation models are widely used because of their ability to capture the dynamic changes in sequential patterns;however,these models treat user interaction records as sequential sequences,but ignore the time interval between interaction items and do not model the actual timestamps.In this paper,we address the above issues by exploring user interaction record data based on the Transformer model,and our main research work is as follows:(1)Markov chains predict the next action based on the user's recent actions and perform well in sparser data sets,while recurrent neural networks predict the items that users are most likely to be interested in next by revealing longer-term semantics in user interaction records and perform better in denser data sets.To balance these two approaches,this paper proposes a sequential model based on the self-attention mechanism that combines information about the items' own features to enable it to capture both long-term semantics and predictions based on relatively few interaction records.With the same recommendation metrics,our model improves the recommendation accuracy by 2% over other models on two datasets(scenic data and Hetrec2011).(2)To address the problem of time intervals between interacting items,this paper models the time stamps in interactions to explore the impact of different time intervals on the next predicted item.To this end,we propose a time-interval-aware sequencebased recommendation model,which models both the absolute position of items in an interaction and the time interval in a sequence.Experiments on two real datasets(scenic data and Hetrec2011)show that the recommendation accuracy of our model improves by 1% over the other best models for the same recommendation metrics.