Research Of Conditional Prefefence Learning And Recommender System

In the age of big data, preference elicitation and recommender system alleviates the problem of information overload. Furthermore, it improves the revenue of online business systems and brings convenience to customers. The research of preference elicitation and recommender system has received greate attentions recent years, which is related to knowledge representation and discovery, decision support, machine learning etc., and is valuable for scientific research. In this thesis, several perspectives about preference elicitation and recommender system are investigated. The main contributions of this thesis are summarized as follows:First, the problem of learning conditional preference netwoks(CP-nets) from inconsistent training samples is addressed. CP-nets, as a simple and intuitive model for representing conditonal preference, has received more and more attention. However, the size of CP-nets increases exponentially with the number of the variables. The computational complexity of the CP-nets learning methods is high and there are many restrict constraints for this kind of methods. Learning CP-nets from inconsistent training samples is difficult especially. In this paper, the problem of learning CP-nets from inconsistent training samples is model and solved. Taking the advantage that dominance testing and consistency testing in preference graphs are easier than those in CP-nets, the proposed CP-nets learning method obtains the CP-nets form inconsistent training samples by two steps. Namely, a preference graph is learn from the inconsistent training samples, and then it is transformed into a CP-net equivalently. It can be proved theoretically that the obtained CP-net entails a subset of examples with maximal sum of weight. Compared with the similar methods, the proposed method get more accurate results on both simulated data and real data.Second, to reduce the computational complexity of CP-nets learning method further, the approximate CP-nets learning method is addressed. Given sufficent training samples, the conditional dependency between two varibales can be judged by hypothesis testing. Accordingly, a CP-nets learning algorithm based on hypothesis testing is proposed. The proposed algorithm is an approximate algorithm, although it do not guarantee to find the optimal results, it can get satisfing results given sufficient training samples. It can be proved theoretically that the obtained CP-net converges in mean to initial CP-net as sample size increases. Furthermore, the proposed algorithm runs in polynomial time, and can be applied widely.Third, how to integrate side information efficiently into recommender system is addressed. In this paper, the model of Bayesian Probabilistic Matrix Factorization (BPMF) is modified. It is assumed assume that the user hyperparameters and item hyperparameters are different for each user vector and item vector. The hyperparameters are generated according to the social relations and item contents. By this idea, a recommender method integrating the social relations and item contents is proposed. The novel way fusing side information is different from traditional regularization-based methods and factorization-based methods. The proposed method can alleviate the data sparsity problem and the cold-start problem better. The proposed methods are computationally efficient and need not any parameter tuning, and it can scale up with respect to very large datasets. Experimental results on three large real world datasets show that our method gets more accurate recommendation results with faster converging speed than the other state-of-the-art recommendation methods based on matrix factorization. Moreover, our method outperforms other methods in cold-start settings. Fourth, how to make the recommendation results more consistent with user preference is addressed. A novel measure is proposed to estimate the differency between the recommendation results and the user preference. A recommender method optimizing this measure, listwise probabilistic matrix factorization (ListPMF), is also proposed Because of the new measure ListPMF gets more satisfing results. Meanwhile, the proposed method can be easily extended to integrate side information, such as social relation and tag information, to improve the performance of the recommender system further. The proposed method is computationally efficient, and can be applied to large-scale real life datasets. Comparison results indicate that our method outperforms other recommender methods based on matrix factorization.Finally, we introduce the future research directions in the preference elicitation and recommender system, including new approixmate contitional preference representation and the recommender method based on conditional preference.