Research On Collaborative Filtering Recommendation Based On Deep Neural Networks

In recent years,it has witnessed the significant success that deep neural networks(DNNs)have achieved in several domains,ranging from speech recognition,computer vision to natural language processing.However,there is relatively little work on personalized recommendation based on DNNs.Although there are some advances employing DNNs to recommendation tasks recently and showing promising results,most of them primarily leveraged DNNs to model side information such as images,visual content,items' textual description.Nevertheless,when it comes to the key to a recommender system-collaborative filtering,they still relied on the traditional matrix factorization(MF)which simply utilized the inner product to model the complex user-item interactions or item-item relationships from user-item interaction data.However,it is unreasonable and insufficient to model highly non-linear real-world data in a linear way.Given the strong non-linear modeling ability to DNNs,we attempt to explore DNNs for collaborative filtering in this work,making use of DNNs to model the complex user-item interactions or higher-order item-item relationships in a non-linear way.In this thesis,we pay our attention to two tasks in personalized recommendation:rating prediction and top-N item recommendation.(1)While MF-based model draw the attention of industry and academia to the rating prediction task for explicit feedback due to its great excellentce,it would expose restrictions on its performance owing to the simple inner product it utilized.To tackle the inherent limitation of MF,we model the potential non-linear interactions between users and items based on the general framework of NCF,short for Neural network-based Collaborative Filtering,so as to address the task of rating prediction for explicit feedback.Extensive experiments on two real-world datasets demonstrate the general framework NCF is capable of modeling the non-linear user-item interactions successfully and achieving encouraging performance in rating prediction task.(2)Compared to MF-based models,item-based CF models have been extensively utilized to building recommender systems in industry due to its efficacy and interpretability.The key to item-based CF models lies in the estimation of item similarities.Early approaches directly utilized vector-space similarity measures to estimate the similarities between items which leads to poor performance since there is no knowledge about item features learned.Although there are some learning-based methods afterwards striving to learn item similarities by constructing an objective function to optimize,they still belong to the family of linear models which limits their performance to some extent.Inspired by the NCF framework,we present a deep learning(DL)view for item-based CF aiming at modeling the higher-order non-linear correlations between items.Meanwhile,we introduce attention mechanism into our proposed model to differentiate the contributions of historical items of a user profile for the final prediction.Comprehensive experiments conducted on two large-scale datasets indicate our proposed approaches can effectively and successfully model the higher-level non-linear item-item relations and outperforms the state-of-the-art methods in the task of top-N item recommendation.