Accuracy And Diversity Oriented Personalized Recommendation Algorithms Research

With the development of Mobile Internet and popularity of Web2.0 applications,recommender system plays a more and more important role in helping users find their interested content and platforms market their products.Among the many optimization goals of recommender systems,accuracy and diversity of recommendation results are the most important,meanwhile they are closely related to each other.Though great interest has been aroused,there are still some problems to be solved:(1)a lack of accurate and general extraction method of user interests and item characteristics which determine the accuracy of recommendation results;(2)In diversifying recommendations,ranking scores of candidate items are usually designed by experience and manually extracted massive features;(3)training deep reinforcement learning based recommendation diversification approaches is instable and low-efficient.To address these problems,in this dissertation,we make the following contributions:(1)A method to leverage multi-modal auxiliary information to help depict user interest and item attributes.By analyzing the content form and description meaning of multiple auxiliary information,we divide them into two types: discrete,and continuous ones.Different methods are proposed to process and use the above information.For discrete ones,we utilize an attention based neural network to dynamically learn their weights,and merge them by adaptive weight to compliment user or item latent vectors.As for continuous content,we adopt pre-trained models by transfer learning to extract their feature representation in high-dimensional vectors,such as VGG16 for image processing,and learn to reduce their dimensions in order to calibrate user and item vectors during training stage.Deploying this method to collaborative filtering shows that the accuracy of recommendation results is effectively improved.(2)A pair-wise learning to rank approach for recommendation diversification,and a design of score function for candidate items based on discounted categories.Different from previous work,the ranking score function is based on widely existing genre information of items,models relevancy and diversity in a unified way.Furthermore,the score function can achieve indirect optimization of non-derivable diversification metric in the training phase due to its design similarity to the evaluation metric.Meanwhile,end-toend learning on the common parameters in the score function avoids the effort of manual feature engineering and its inaccuracy.Experiments demonstrate that diversity of recommendation results with our proposed method performs far better than those comparative methods.(3)An data efficient Actor-Critic reinforcement learning algorithm for diversifying recommendations.We formalize the recommendation diversification problem into an MDP setting,and adopt LSTM to model user's diversified interest requirement state changes with the recommending items and calculate the selection probability distribution of candidate items.Different from existing policy gradient based methods or typical Actor-Critic algorithms,we leverage Bellman Expectation Equation and expected policy gradient formula in whole action space to estimate the update gradients of both Actor and Critic parameters based on one-step forward ahead.Theoretically,using full actions at a state can improve accuracy and data efficiency.Experiments results achieve better diversity compared to multiple baselines,and outstanding training stability and efficiency over other reinforcement learning based approaches.