Research On Recommendation Strategies Under Uncertainty Information

As an efficient information filtering technique,recommender systems(RSs)can filter out the content that users may be interested in according to their needs for personalized recommendations.Meanwhile,RS,as a specific uncertainty information system,contains a variety of uncertainty information,such as fuzzy review information and inaccurate rating information.If the uncertainty information is not addressed effectively,the accuracy of recommendation results will be seriously affected.Therefore,under the condition of uncertain information,various recommendation solutions are designed to solve the following key problems by using uncertainty mathematical knowledge: 1.Inconsistency between user ratings and reviews and natural noise;2.Data sparsity problem caused by relying on the co-rated items;3.Trade-off between accuracy and efficiency and cold-start problem;4.Scalability problem.The main contributions of this dissertation are as follows:(1)Regarding the problem that most rating-based collaborative filtering(CF)recommendation methods only use users’ original rating information,a sentiment based multi-index integrated scoring method is proposed to pre-process and fuse a variety of uncertainty information to provide reliable information input that can reflect users’ comprehensive preferences for recommendation algorithms.Firstly,a sentiment analysis based expanded lexicon method is used to transform the reviews into normalized sentiment scores.Then,a multi-criteria classification based natural noise detection method is exploited to identify and correct the incorrect ratings.Finally,a weighted average method is employed to integrate reviews and ratings and obtain comprehensive ratings.Compared with only using original ratings,CF methods using comprehensive ratings can obtain better accuracy of recommendation results.(2)To solve the problem that memory-based CF methods mainly often rely on the co-rated items,a K-medoids clustering recommendation algorithm based on probability distribution is proposed to make full use of all rating information.Firstly,a probability method is adopted to calculate users’ preference probabilities,and the KL divergence in information theory is used to fully measure item similarity.Then,a K-medoids clustering algorithm based on KL similarity is introduced to classify items,so that the nearest neighbor set can be found quickly and accurately to predict ratings.The experimental results on public datasets indicate that our model is superior to the comparison methods in terms of all metrics,which alleviates the data sparsity problem to a certain extent.(3)For the problem that a single memory-based CF method deals with user rating information under different number of co-rated items,a hybrid similarity recommendation method through integrating multiple information is proposed to realize the trade-off between accuracy and efficiency.Firstly,an adjusted Google similarity method is introduced to calculate semantic similarity between items quickly and accurately in the condition of enough co-rated items.Subsequently,an intuitionistic fuzzy set based KL similarity is presented from the perspective of user preference probability to measure fuzzy similarity between items in the condition of rare co-rated items.Finally,the two similarity methods are integrated by an adjusted variable to comprehensively evaluate similarity values when the number of co-rated items lies in a certain range of values.The experimental results on real datasets indicate that the proposed method has higher accuracy and lower time complexity than the comparison methods,which has a positive effect on addressing the cold-start problem.(4)According to the problem of matrix factorization(MF)based CF models mainly using explicit rating information,a probabilistic matrix factorization(PMF)recommendation model for integrating multiple information is proposed to provide reliable predictions and recommendations.Firstly,a sentiment analysis based PMF model is introduced to fit normalized sentiment scores.Then,the proposed natural noise detection method is adopted to transform the rating matrix into a binary reliability matrix,a reliability based PMF using Bernoulli distribution is presented to factorize the binary matrix.To effectively integrate three types of information into a PMF procedure,a weighted matrix with a uniform weighting strategy to obtain predicted ratings with the corresponding reliability probability.The experimental results show that the proposed model is superior to other MF methods in terms of accuracy and scalability.Overall,this dissertation takes uncertainty information sources as an entry point and applies uncertainty mathematical methods to handle user information and build recommendation models,which provides new research insights for improving the accuracy of RSs.In future work,some related solutions based on current studies can be designed to enhance the diversity of RSs;thus promoting the recommendations of long-tail items.