Research On Recommendation Algorithm For Accurate Expression Of User's Personalized Interest

At present,there are recommendation systems everywhere on the Internet,such as e-commerce,news,short video and so on.In the case of information overload,the recommendation system can quickly screen out the goods that users may be interested in from the candidate set.It is not only a sharp tool for users to obtain information efficiently,but also can improve the profits of businesses.Recommender system has become one of the core technologies in Internet applications,and it is also a powerful engine to promote the growth of Internet.In this paper,some new ideas and algorithms are proposed based on some common practical problems in the application of recommender system.The details are as follows:(1)Firstly,in the recommendation system,an embedding vector is often calculated for each user to represent the user.The current calculation methods can hardly avoid emphasizing some behavior features of users,which results in the cover up of user basic features,and then makes the user embedding vector deviate.This paper proposes a two tower model to isolate user attributes and behavior.Two neural networks are used to process user attributes and behavior features respectively,so as to avoid the covering effect of behavior features.The output of the two networks is connected to get user embedding.You can also easily control the dimensions of attribute vector and behavior vector to adjust the weight of user behavior.In order to verify the effect,this paper carries out simulation experiments on the international public data set,and compares with the common recall methods.The experiments show that the user double tower embedding model has better performance.(2)Secondly,in the relevant recommendation scenario,that is,to continue to recommend related items under a main item page,the current solutions are not comprehensive enough to measure the user's interest in the main item.This paper proposes the DCNet model to express it explicitly.The directional crossing layer is used to complete the crossing in the case of linear complexity,which improves the efficiency to the greatest extent.According to the characteristics of the wide part and the deep part,different crossover methods are used to ensure that the model can learn the user's interests most accurately.This paper tests on the real Taobao dataset published by Alibaba,and compares it with the current mainstream click through rate prediction model.The results show that DCNet model can better express users' interests and has better performance.(3)Finally,in the recommendation scenario,the user's behavior is related to the time sequence.A certain user's behavior is not only affected by the historical behavior,but also makes the future interest change.The current mainstream recommendation algorithms regard each user request as an independent calculation,which does not meet the actual needs of the scene.This paper proposes a set of solutions to the obstacles of user serialization behavior modeling and reinforcement learning.Firstly,the simulated user is trained by historical data,which is used as the environment to train the model.Item embedding is used as an action expression in the model.Naturally,nosy net is introduced to explore actions.Using the BERT+LSTM model to extract the user's historical behavior and obtain the user status.The DDPG model is used in the simulation experiment,and the results show that it has better performance than the supervised learning model.