Research On Key Technologies And Applications For Machine Learning Based Recommender Systems

In recent years,Recommendation Systems(RS)have played a vital role in information filtering and retrieval,and have been widely applied in such social networks,ecommerce,news recommendation and other fields,and have achieved huge economic and social benefits.However,RS still suffer from such as data sparsity and cold start problems,which could lead to low recommendation accuracy.Therefore,how to solve the information overload problem and provide users with high-quality recommendation information is still a current research hotspot,which has important research value.This dissertation generally focuses on the key technologies and applications of machine learning based recommendation systems,which could solve the problem of information overload,and provide users with effective and efficient personalized recommendation.The research content generally contains four parts.(1)Based on pairwise learning to rank method,this dissertation generally focuses on user's preference exploitation via taking account of the implicit feedback,and proposes a recommender model referred to as BPLR.Based on the model for user's interest exploitation,BPLR divides items into positive feedback,potential feedback,and negative feedback,according to the user's historical behaviors,which could be helpful for fine-grained interests exploitation.The collection of items with potential feedback is derived from the user's social network and item popularity,which could guarantee the accuracy and completeness of recommendation results.BPLR employs Bayesian inference for model inference and optimization,and uses Stochastic Gradient Descent(SGD)for parameter learning.In addition,to filter out invalid training samples and select informative samples,the dissertation proposes a dynamic sampling strategy(DSS),which could reduce computational complexity and speed up the model training.Implicit feedback is easy to collect in real life applications,and it's promising for BPLR to be applied to utilize the implicit feedback for recommendation.In addition,this dissertation also provides theoretical convergence analysis for BPLR,including the corresponding theorems and proofs.(2)This dissertation focuses on the impacts of users' social network and interests on the recommendation results,and proposes a recommendation model called SRMP.This model further subdivides each user's trust relationship into trust and trusted relationship,and calculates the expert level value for each user in the social network as the user's weight,which is to constrain the low-rank feature vectors learning for users and items in matrix decomposition.In the model optimization phase,SRMP classifies items into multi-categories according to category labels,and treats items in each category as an independent social community,after that,SRMP performs model training and score prediction in each social community,to improve the recommendation accuracy.It's already certificated that SRMP could capture the intricate social relationship between users,and learn each user's interests,which could alleviate the data sparsity of the Recommendation Systems.In addition,SRMP could also utilize the social relationship to solve the cold start problem,and significantly improve the recommendation accuracy and user experience.(3)The deep learning methods has powerful capability in information processing and retrieval.This dissertation focuses on the principles and technologies of deep learning based recommender algorithms,which take account of the numerical ratings,user portraits,item attributes,and available additional information to provide accurate recommendation,and proposes a deep learning based recommender model called DLMR.This model generally includes recommender candidates generation and candidates ranking.In the phase of candidates generation,DLMR employs the topic distribution model LDA to learn the user's interests distribution from the textual reviews and item's contextual information,and then integrates it into the Convolution Matrix Factorization(CMF)model for feature vectors learning for users and items,after that,DLMR could perform rating prediction to generate the recommender candidates.In order to further improve the recommendation accuracy,DLMR employs a three-layer Denoising Autoencoder(DAE)network,which could utilize available Auxiliary Information(AI)to rank the generated recommender candidates,to generate the final top-N recommendation list.In real life applications,DLMR could capture the complex relationship between users and items,exploit the interest distribution for users,and provide users with accurate personalized recommendation.Therefore,deep learning based recommender algorithms have powerful capability in data processing,which could effectively utilize available information with various kinds of attributes to improve the accuracy of recommendation results,and it's promising for deep learning based recommender systems in various real life applications.(4)This dissertation focuses on the data sparsity in large-scale online recommendation,and the problem of online update,and proposes an online recommender model referred to as LsRec,which generally includes offline calculation and online update.LsRec calculates each user's social influence according to the social network,and then integrates it into the matrix factorization model for feature vector learning for users and items.In order to provide accurate recommender results,LsRec employs the TF-IDF method to represent items in the form of vectors,and then performs item clustering via k-means++.After that,LsRec performs model learning in each generated item clusters respectively,to improve the recommender accuracy.In order to ensure the practicability and convergence of the model,the dissertation provides some theorems and provides the corresponding theoretical proofs.In order to address the problems of online update and cold start problem,based on the fast matrix transformation method,LsRec proposes an online incremental update strategy,which could timely capture the evolution of users' interests,scale flexibly over large-scale datasets,and provides accurate recommender services for users.In largescale online recommendation applications,LsRec could effectively alleviate the problem of data sparsity,perform online update quickly and efficiently,and provide users with timely and efficient online recommendation service.