【作者】 陈钰；

【导师】 彭强；

【作者基本信息】 西南交通大学 ， 计算机科学与技术， 2017， 硕士

【摘要】 信息技术和互联网的迅猛发展把人们带入了信息时代,信息量增长迅速是这一时代的显著特征。面对海量数据,用户往往不能快速而有效地从这些信息中获得对自己有用的信息,这就是大数据时代的“信息过载”问题。如何从大量信息中过滤出有价值的信息已经成为业内的研究热点和亟待解决的问题。搜索引擎和推荐系统的出现在一定程度上解决了 “信息过载”的问题。面对互联网上日益增长的信息,被动地应答用户查询的搜索引擎已经不能完全满足广大用户的需求。与搜索引擎的被动式应答方式不同,推荐系统可以根据用户的兴趣、行为、情景等信息,通过相关的推荐算法把用户最可能感兴趣的内容主动地推荐给用户,因此得到了广泛的研究并取得了长足的发展。协同过滤推荐算法在推荐系统的发展过程中起到了重要的作用,这种算法在推荐系统领域得到了深入研究和广泛使用。但传统协同过滤推荐算法面临着难以避免的数据稀疏及冷启动等问题,这些问题严重影响了推荐的质量,如何有效解决传统协同过滤算法面临的问题成为了业内的研究热点。近年来,随着社交网络的蓬勃发展,大量的社交信息数据也随之产生,因此有学者提出应该充分挖掘社交信息的潜在价值,利用社交信息来改善推荐的效果。信任信息作为社交信息中的重要表现形式被引入到推荐算法中,许多基于信任的推荐算法也被相继提出,这些算法从不同的角度缓解了数据稀疏和冷启动问题,改善了推荐的效果,但这些算法也存在一定的问题,本文就进一步改善推荐算法的效果进行了以下两方面的工作。首先,针对推荐系统的评分预测任务,在SVD++模型和TrustMF模型的基础上研究了一种新的评分预测模型TMFSVD,该模型充分利用了 SVD++模型和TrustMF模型的优点,通过矩阵分解技术将评分信息和用户信任关系联系起来,综合考虑了评分信息的显隐式反馈和用户信任关系对预测评分的影响。在3个被广泛使用的真实数据集上的实验表明,TMFSVD模型可以有效地缓解数据稀疏和冷启动问题,且能提高评分预测的准确度。其次,针对推荐系统的Top-N推荐任务,研究了一种基于信任关系的Top-N推荐模型TrecRank,该模型通过矩阵分解技术将用户评分信息和用户信任关系联系起来,综合考虑了用户评分信息的显隐式反馈和用户信任关系对Top-N推荐的影响。TrecRank模型通过Plackett-Luce模型来对预测列表进行排序,并根据信任网络中节点的出度和入度的数量对用户间的信任值进行了修正。在3个被广泛使用的真实数据集上的实验结果表明TrecRank模型可以有效提高Top-N推荐的准确率。更多还原

【Abstract】 The rapid development of information technology and the Internet has brought people into the information age, the rapid growth of information is a significant feature of this era. In the face of a massive amount of data, users frequently can not obtain information that is useful to them quickly and efficiently, which is the "information overload" problem in the era of big data. How to filter out valuable information from a large amount of information has become a research hotspot in related fields and an urgent problem which needs to be solved. The emergence of search engines and recommendation systems has solved the problem of"information overload" to a certain extent. In the face of growing information on the Internet,search engines that passively answer user queries can no longer fully meet the needs of the vast numbers of users. Unlike the passive response mode of the search engine, recommendation system can actively recommend the items that users may most interest in to the users according to the user’s interest, behavior, scenario and other relevant information through some relevant recommendation algorithm, so it has been widely studied and achieved great development.Collaborative filtering recommendation algorithm plays an important role in the developing process of the recommended system, and it has been intensively studied and widely used in the field of recommendation system. But traditional collaborative filtering recommendation algorithms are faced with such inevitable problems as data sparseness and cold-start, these problems seriously affect the quality of recommendation, how to solve the problems of traditional collaborative filtering algorithm effectively has become the focus of research in the correlation domain. In recent years, with the rapid development of social networks, a large number of social information data also has been generating, therefore,some researchers suggest that we should fully exploit the potential value of social information,so that we can use it to improve recommendation algorithm and make it better. As an important form of social information, trust information has been introduced into recommendation algorithms, and many trust-based recommendation algorithms have been proposed. These algorithms alleviate the data sparseness and cold-start problems at a different angle and improve the recommendation effect. However, there are some problems in these algorithms. In this paper, the following two aspects of work are done to further improve the effect of the recommendation algorithm.Firstly, for the rating prediction task of recommender systems, this article researches a novel rating prediction model, termed TMFSVD, this model is based on the the SVD++ model and TrustMF model, TMFSVD model makes the best of the advantages of SVD++ model and the TrustMF model,TMFSVD model bridges ratings information and user trust relationships together through matrix factorization technique and incorporates both the explicit and implicit influence of user-item ratings and user trust relationship on the prediction of items for an active user. Experimental results on the three widely used real data sets demonstrate that TMFSVD model can reduce the data sparsity and cold start problems effectively, and it can improve the accuracy of rating prediction.Secondly, for the Top-N recommendation task of recommender systems, this article researches a novel Top-N recommendation model, termed TrecRank, which based on the user trust relationships. TrecRank model bridges ratings information and user trust relationships together through matrix factorization technique and incorporates both the explicit and implicit influence of user-item ratings and user trust relationship on the Top-N recommendation of items for an active user. TrecRank model ranks the items in the predicted list through Plackett-Luce model, and adjusts the trust values between two users according to the number of outdegrees and indergrees of nodes in trust network. Experimental results on the three widely used real data sets demonstrate that TrecRank model can improve the accuracy of Top-N recommendation.更多还原