The Model Based On Heterogeneous Behaviors In Sessions For Click-through Rate Prediction

Recently,the continuous growth of information has become one of the factors perplexing people’s life,and the emergence of recommendation systems has brought convenience to people’s life to a great extent.As an important part of the recommendation field,the hit rate prediction task also has important significance.The improvement of click through rate can not only provide a basis for users to quickly screen information and products with their preferences,but also bring huge benefits to the company.Many factors are involved in the process of click through rate prediction,and relevant researchers are constantly exploring the impact of these factors on the click through rate prediction task.The existing research methods of click through rate prediction have the following shortcomings: first,there are a variety of heterogeneous behaviors in the user’s behavior sequence,which can show the user’s interest to varying degrees.The existing research does not fully consider the impact of different behaviors in the user’s behavior sequence on their interest;Second,there may be multiple interests in the same session.The existing studies assume that a session corresponds to only one interest,and can not distinguish the dependencies between specific types of interests.In order to solve the above problems,this paper does the following research:Firstly,in view of the insufficient consideration of the impact of user interest on their behavior types,the Model Based on Heterogeneous Behaviors in Sessions for Click-through Rate Prediction(HBS4CTR)is proposed,which processes the behavior sequence generated when users interact with items according to the behavior types.The attention mechanism is injected into different behavior types of the same session to learn the influence weight of each type of behavior on session interest.And the model forms a new session representation,then extract the session interest.Finally the model learns the relationship between the session interest and the target item,and predicts the click through rate.Secondly,aiming at the problem of multiple interests in the same session,in order to avoid ignoring the user interests represented by a small number of interactions after mixing a large number of interactions of the same type and a small number of interactions of another type,and improve the accuracy of user interest extraction,a method of separating session interests is proposed to reasonably segment and recombine the behavior sequence generated when users interact with items.For each interaction item type of the session,a corresponding interest is extracted.Finally,in order to verify the performance of HBS4 CTR model,this paper compares the proposed HBS4 CTR model with You Tube Net,Wide&Deep,DIN,DIEN and DSIN on Taobao advertising click / display dataset.The experimental results show that the HBS4 CTR model proposed in this paper has better effect in predicting click through rate.