Research On Personalized Recommendation Methods Based On Deep Learning

Recommendation is an important technology to solve information overload in Internet platforms such as shopping,education,news and entertainment.It is a research hotspot in the fields of information retrieval and data mining.The technology has been successfully applied to personalized recommendation of products,content and services on various Internet platforms.It cannot only maximize user preferences and help users obtain information of interest,but also achieve the business goals of information platforms while maintaining the good user experiences.With the development of deep learning and natural language processing,the recommendation field has formed a hybrid recommendation method based on machine learning.Compared with traditional collaborative filtering-based recommendation methods and content-based recommendation methods,this kind of hybrid recommendation method can mine sequence information from multi-source heterogeneous data,and automatically extract rich hidden features of users and items,which can more effectively reflect the different preferences of users to achieve personalized recommendation.This kind of method has been paid great attention by academia and industry,and will be more widely and deeply studied and applied in the future.The current hybrid recommendation methods based on deep learning solve some problems of traditional recommendation methods to a certain extent(among them,the problems include the dependence on manual feature extraction and the difficulty in extracting features from multi-source heterogeneous data).However,they ignore the rich hidden features which are contained in the combined information and overall information in structured data,and also ignore the problems of word sparseness and word synonymy in text data(i.e.unstructured data).In addition,the existing hybrid recommendation methods based on deep learning are supervised learning methods that highly rely on labeled data.These methods are difficult to apply to extreme environments such as new systems,new fields,or high levels of privacy protection,because it is difficult to collect usable manual annotation data.Therefore,solving the cold start problem in harsh environment is an urgent scientific problem in the field of recommendation,which has important research significance.In response to this urgent and challenging scientific problem,this dissertation conducts in-depth research on attention mechanism,prioritized experience replay mechanism,and adaptive improvement of neural networks.To be specific,the major research work of this dissertation is summarized as follows:1.In view of the data sparseness,cold-start problem and excessive dependence on manual feature extraction in traditional recommendation methods,this dissertation proposes a novel recommender with attention-based convolutional neural network and factorization machines.Fir-stly,we propose a word-level attention mechanism and a phraselevel attention mechanism to form a "word-phrase" attention mechanism based on natural language processing,which explores the contributions of core vocabulary and important phrases in comments of users and items from a local to a global perspective.Secondly,we establish a two-column Convolutional Neural Network(CNN)model based on the hierarchical attention mechanism,which can mine the hidden expression features of users and items.Finally,we use the factorization machine to mine the correlations from their hidden features to form an efficient hybrid recommendation approach(i.e.ACNN-FM).2.Aiming at the problems of data sparseness,cold-start problem,and difficulty in mining valuable long-tail POIs for individual users in personalized Point-Of-Interest(POI)recommendations with massive check-in data scale,this dissertation proposes a personalized POI Recommendation based on Hierarchical Attention Mechanism(HAMPOIRec)that can effectively improve data utilization.Firstly,in order to mine more hidden features from limited data,we propose the concepts of explicit features and implicit features,which can guide POI recommendation methods based on deep learning to provide ideas for selecting data and computing models.Secondly,we propose a "localoverall" structure of attention mechanism,which can locally focus on the contribution of a single feature to POI recommendation,and mine the contribution and hidden features from the combined features and overall feature to POI recommendation as a whole.Finally,in the field of POI recommendation,we propose a "user-POI" matching mechanism for the first time based on natural language processing,and use this mechanism to fine tune the user's POI recommendation list to achieve higher precision recommendation.3.In order to solve the problems(such as excessively large and dynamically variable of users and items,highly discrete action space,and extremely sparse environment interaction feedback)in Deep Reinforcement Learning(DRL)recommendation field,and to implement an reinforcement learning-based recommendation method that does not depend on manual annotation data,this dissertation presents a Deep Reinforcement Learning Recommendation(called HEDRL-Rec)based on Hierarchical attention and Enhanced priority experience replay mechanism.Firstly,we present an Actor neural network based on hierarchical attention mechanism,which can mine the auxiliary information contained in user(or item)status and its contribution to recommendation from the local and global features(including composite features).Secondly,we propose an enhanced priority experience replay mechanism that reuses both historical experience and considers the different importance of experience,which can alleviate model overfitting while resolving the problems of sample imbalance,difficult convergence and low learning efficiency.Finally,we propose a deep reinforcement recommendation training method based on Deep Deterministic Policy Gradient(DDPG)architecture,which effectively guarantees the convergence and fitability of the model,alleviates the cold start problem in the field of recommendation,and realizes personalized recommendation independent of manual annotation data.