Research On Deep Neural Network Recommendation Algorithm Based On Attention Mechanism

In recent years,the recommendation system has developed rapidly,and has received great attention in industry and academia.At present,there are many solutions for recommendation systems,the more traditional and classic is the collaborative filtering recommendation algorithm.Model-based recommendation algorithms are also a big category.Among them,the attention mechanism method can focus on the features that best reflect the user's interest preferences,and can better capture the dynamics and diversity of user interests,so based on attention The mechanism of recommendation algorithm research is of great significance.In order to prove that the attention mechanism has a better effect than DNN and RNN when dealing with short-term sessions,first of all,a recommendation algorithm based on short-term sessions is studied.Recommendations are made based on the user's short-term behavior characteristic information,and position coding is used to obtain user behavior data.Relative position information,then capture the short-term interest of users through the multi-head self-attention network,and then convert it into a single embedding vector through the attention mechanism,and finally enter the DNN model for prediction and recall steps.Considering the user's long-term behavior characteristic information,in order to make the recommendation algorithm have better effect,on the basis of short-term session,merge long-term session data,further study a recommendation algorithm based on long and short-term session,use multiple attention in parallel to capture the historical behavior features of multiple users in the long-term session,and then use the attention mechanism to assign weights to the long-term and short-term features to construct the final embedding vector.Experiment with the two models studied,and add the traditional collaborative filtering model and some models based on deep learning as comparative experiments,and evaluatethe model through the four indicators of recall rate,accuracy rate,F1 Score and Hit Ratio.The analysis of the experimental results found that the recommendation model based on short-term session has good performance on various indicators,while the recommendation model based on long-short session has the best performance,and the Hit Ratio is improved by 6.83% compared with the benchmark model.Compared with the DNN-based recommendation model The recommendation algorithm is improved by 5.9%,and the experimental results verify the feasibility and effectiveness of the designed model.At the same time,the parameters of the model are analyzed experimentally.The results show that the model performs best when the number of multi-head self-attention networks is 4 and the interest fusion method is attention.