Research On Recommendation Model Combining Ladder Idea And Deep Neural Network

In recent years,the exponential growth of Internet information,data becomes more and more large,users are faced with more and more choices,but everyone has different interests and hobbies,in the face of such a variety of choices how to accurately locate their preferences has become the key issue.With the emergence of various recommendation algorithms,this problem has been alleviated,but it ignores the potential connection between the user and the project in the recommendation process.In recommendation algorithm,users and items are often not independent,and there are a lot of potential features between them.The traditional recommendation algorithm only predicts the items users may like based on the interaction information between users and projects,ignoring the potential correlations between different users and different projects,which is not enough to deal with the cold start problem brought by a large number of new users.In terms of feature extraction,the potential features extracted by the existing mainstream models only exist in high and low order,and these feature dimensions are too single to fully explore the potential preferences of users for the project.The main research contents of this paper are as follows:(1)Users and items are introduced into the model in pairs,and the double similarity between different users and different items is calculated.The attention mechanism is used to learn the different preferences of users for items and the correlation between items and users,so that the model can fully learn the potential preferences of users.(2)Improve the traditional CNN into a stepped CNN,extract the potential features of users and projects from the three dimensions of low order,middle order and high order,assign appropriate weights to the three feature channels through the attention mechanism,and combine the three feature channels in pairs according to the weights learned by the model.Three new feature channels,low-middle-order,low-high-order and medium-highorder,are formed.Finally,six feature channels are fused into the fully connected network layer in parallel to complete the final recommendation.The above improvements not only retain the functions of CNN convolution layer and pooling layer,but also realize multidimensional extraction of potential features between users and projects,which solves the problem of insufficient dimension of feature extraction in current mainstream algorithms.(3)Based on the original data set,the auxiliary information of users and projects is introduced.The feature extraction by stepped CNN enables the model to complete accurate recommendation according to the auxiliary information of users and projects without user interaction information,which improves the recommendation accuracy and ranking performance of the model and better alleviates the cold start problem.Finally,the two models proposed in this paper are jointly modeled,and a representative model in the field of recommendation algorithms is selected as the baseline model.The experiment of two representative public data sets in the field of recommendation algorithms proves that the model proposed in this paper has higher recommendation accuracy and ranking performance,and at the same time better alleviates the cold start problem.