Research On Item Collaborative Filtering Algorithm Based On Historical User Information

With the explosive development of Internet data,people's access to information services has changed from nothing to better habits.At the same time,in the face of the Internet's increasingly complex items,it has become dazzled and unable to make decisions.When faced with a data load,if the user does not know how to express or cannot describe things,can he recommend related items based on past interaction behavior records? The recommendation system can help users quickly obtain the required information.In reality,the interaction data between users and items is often sparse,which brings certain difficulties to the recommendation algorithm.The core of the recommendation system is to model the user's information according to the user's historical behavior preferences,item attributes or context information,so as to achieve an effective recommendation function.The most widely used algorithm in traditional recommendation systems is collaborative filtering,which uses the interaction between users and items to generate recommendations.The collaborative filtering-based recommendation method uses a shallow model that cannot learn the deep interaction features of the user and the item.The learned features are linear.With the advent of deep learning,it brings new opportunities for traditional recommendation systems.On the one hand,deep learning converts the learning relationship of the traditional recommendation system from linear to non-linear,which can obtain the representation of non-linear features between users and items;on the other hand,deep learning can automatically learn from data features and map different data In the same space,this greatly solves the problem of data sparsity,thereby alleviating the problem of cold start and sparse data in the recommendation system.However,in the deep model,more meaningless neural factors are generated,which cannot explain the recommendation results reasonably.At the same time,extracting features is an automatic learning process,and it is impossible to distinguish the importance of features.Therefore,this paper mines the information of historical users of the item,effectively alleviates the problem of data sparseness to a certain extent,and integrates the network of attention mechanisms to distinguish the importance of different users of the item.At the same time,on the basis of historical users,construct the historical sequence of historical users,model the sequence information of users and items,and mine the long-term and short-term intentions of historical users.Sequence recommendation uses context information to model user behavior sequences so that the recommendation system can truly perceive past behaviors and influence future interactive behaviors,and provide an interpretability for the recommendation results.The main work of this dissertation includes the following:1.This dissertation first introduces the background knowledge and research significance of the information load of the recommendation system,fully investigates the sparse data affects the performance of the recommendation results and the recommendation results generated by deep learning cannot explain the research status and solutions of such problems,and then analyzes the user-based The challenges and difficulties faced by collaborative filtering algorithms,and specific research solutions are given for these challenges.2.In view of the shortcomings of the traditional user-based collaborative filtering recommendation algorithm and the importance of relying on data characteristics in the deep learning process to effectively distinguish data,a collaborative filtering recommendation algorithm based on historical user similarity is proposed.The algorithm focuses on modeling the interaction between historical users and items,and learns the similarity matrix between users from historical user interaction data.At the same time,it introduces a network of attention mechanisms to distinguish the importance of different users to the item.3.Traditional recommendation systems rarely consider that time has an important influence on user behavior(or future behavior),and time will determine the user's intention in the long-term or short-term.To solve this problem,a recommendation algorithm based on long-term and short-term intentions of historical users is proposed.Based on historical users,the algorithm introduces time series to mine long-term or short-term intentions of historical users,effectively learns user-item interaction features,and finally merges two features to learn the deep-level features of users and items.