Research On Personalized Recommendation Technology Based On Representation Learning

With the rapid development of the network,the information resources are increasing day by day,and the problem of information overload occurs.In the face of massive data,users cannot explore the information they are interested in,and it is difficult for businesses to display information to appropriate users.Recommendation systems can effectively alleviate the problem of information overload and provide users with personalized recommendations.At present,there are still some problems in the recommendation algorithm,such as sparse data,insufficient information utilization,high model complexity,and low prediction accuracy.The data determines the upper limit of the algorithm model.In view of the above problems,this thesis uses representation learning to automatically learn the effective features of the data to obtain more efficient and meaningful vector representations,and combines representation learning to propose two recommendation models,which are tested on real data sets and systems.performance.The main research contents of this thesis are as follows:(1)Aiming at the data sparsity problem and the lack of correlation between encodings when traditional recommendation algorithms use one-hot encoding and Hash encoding to represent users and items,a collaborative recommendation model Profile DNN based on word2 vec deep neural network is proposed.The model learns the item vector from the user interaction history through word2 vec,which is low-dimensional and dense and contains similar information between items.Then combine the item vector and user preferences to build a portrait for the user,and learn the high-level feature interaction between the user and the item through the deep neural network,so as to achieve rating prediction and recommend high-rated items for the user.(2)The Ebbinghaus forgetting curve reveals the law of memory decay over time,indicating that maintaining long-term memory requires constant review.This thesis takes full advantage of the Ebbinghaus forgetting curve and proposes a multi-task matrix factorization recommendation algorithm TAMMF based on temporal attention.The model retains the reproduction recommendation method of matrix factorization-vector inner product,adopts the attention mechanism to capture the adjacent information of users and items,considers the characteristics of user preferences changing with time,and uses the Ebbinghaus forgetting curve to describe the time decay characteristics of adjacent information.During the training process,the model introduces the experience playback method of reinforcement learning to repeat the training to simulate the review process during memory.The final training obtains the vector representation of users and items,which is easy to reproduce the recommendation and model deployment,and has a high prediction accuracy.(3)Using TAMMF as the core recommendation algorithm,build a movie recommendation system based on B/S architecture.The system is divided into display layer,logic layer and data layer,which are respectively responsible for user registration and login,movie rating,personalized recommendation and other functions.The system is developed based on Spring Boot,My SQL,Mybatis-plus,Redis,etc.,and JMeter is used to stress the system.