Research On Sequence Recommendation Algorithm Based On Deep Learning

The recommendation systems can analyze and model the historical behavior of users and generate recommendations based on the learned user interests,reduce the negative impact of information overload and improve user experience.Generally,the user behaviors usually occur sequentially and there are context dependencies in the front and back interactions.Sequential recommendations learn the representations of user interests by modeling the temporal dependencies in the interaction sequence and predict the next interaction behavior of users.Although sequential recommendations have been widely studied and applied,the existing methods still have the following problems: first,the dynamic change characteristics of short-term interests and the global stability characteristics of long-term interests are not considered at the same time.When learning user interests,complex dependencies and transitions between items in short-term sequences are not adequately modeled,the importance of different item features is not distinguished according to user preferences.Second,the interaction time between the user and the item is not considered in the modeling,they are only based on the sequence of interaction,long intervals will lead to less relatedness between items.For sequences with different time spans,the focus of modeling should be different.Third,the user interests are only modeled according to the interaction behaviors in a single sequence,they do not consider the important item collaboration information between sequences.When the data in a given sequence is too sparse,the model can not make accurate prediction.In order to improve the accuracy of recommendation,this paper mainly studies the above problems from the following three aspects:(1)Sequential recommendation algorithm based on self-attention mechanism.In the short-term interest modeling,the self-attention mechanism can calculate the importance of other items to the current item in parallel,it measures the correlation between different items.In long-term interest modeling,a user-based gating network is used to extract the features of items that users focus on.Finally,short-term interests and long-term interests are combined to predict the user's next interaction behavior.Experiments show that the method proposed in this paper achieves the best recommendation effect under the two evaluation indicators.(2)Sequential recommendation algorithm based on temporal self-attention and multi-preference learning.In dynamic interest modeling,the intention representations of users at each time step are obtained through GRU,the temporal gate self-attention mechanism is used to capture the dynamic changes of user intentions,the drift process of user interests is modeled with the help of time information,it can improve accuracy.When modeling general interests,the multi-preference matrix is used to classify user's preferences,it can increase diversity.It recommends items for users by fusing users' dynamic interests and general interests.The experimental results show that the method proposed in this paper significantly outperforms the existing methods on the tasks of sequential recommendation.(3)Sequential recommendation algorithm based on graph neural network.Firstly,an item collaborative graph is constructed for the current interaction sequence.The improved graph attention network is used to aggregate the item information of the neighborhood,both the co-occurrence relationship between the two items and the time constraint relationship are considered.Temporal gating is used to address the effects of time interval on item relevance when modeling user dynamic interests.When modeling user static interests,user-based attention mechanism can generate personalized static interest representations for different users.The method proposed in this paper can combine and use the item information in the sequence and the collaboration information between sequences at the same time.The experimental results show that it can effectively reduce the impact of data sparsity and provide more accurate recommendations.