个性化推荐系统(Personalized Recommender System)是近年来兴起的热门研究领域,它是一种有效解决信息过载问题的工具。虽然研究人员提出了很多相关的方法来推进推荐系统研究,但是个性化推荐技术依然面临很多挑战,如评分预测准确度、数据稀疏性、冷启动等问题。而另一方面,新近兴起的在线社交网络(Online Social Networks)提供了大量的用户行为数据和关系数据,如何有效的利用这些社交网络数据提升传统推荐算法的精度,以及解决新用户冷启动问题是本文关注的重点。 本文针对社会化推荐系统中的评分预测问题和Top-N推荐问题进行深入研究,主要工作和贡献包括以下几个方面: 1) 针对现有的社会化推荐算法通用性和可扩展性差等问题,在现有的基于矩阵分解的社会化推荐算法和Factorization Machine(FM)模型的基础上,本文提出一种运用特征工程方法,将各种社交网络信息融入到传统矩阵分解模型中。同时还提出一种基于特征组交叉的方法(FM-GI)改进传统的FM模型。在真实数据集上的实验结果表明:在评分预测问题上,该方法能有效的提升评分预测的精度,减少计算复杂度,并且能部分解决新用户冷启动问题。 2) 推荐系统的最终目标是给用户提供一个排序的推荐列表,本文提出一种基于排序学习(Learning to Rank;LTR)的方法来融合基于社交网络信息和传统协同过滤的混合推荐算法LTR-HCF,该方法能有效的将各种社交网络信息和传统的基于邻域的协同过滤算法集成,并且具有很好的可扩展性,实验结果表明:在Top-N推荐中,和先前的方法相比,该方法在有效的提升推荐精度的同时,能部分解决新用户冷启动问题。
英文摘要
In recent years, recommender systems have been proposed as a key tool to overcome information overload, and many algorithms and systems have been developed. Despite all these efforts, recommender systems still face many challenges—such as improving prediction accuracy, data sparsity, and cold-start issues. On the other hand, the recent emergence of online social networks provide us with an unprecedented large amount of information regarding user behavior and friend interactions. This paper focus on how to improve traditional recommender system effectively with social network information. In this dissertation we investigate the rating prediction task and Top-N recommendation task in social recommendation by utilizing information in social networks. The main work and contribution of this dissertation are listed as follows: 1) For the problem of generality and scalability in existing social recommendation algorithms,based on the matrix factorized social recommendation and Factorization machine, we proposed a feature engineering method to incorporate social network information into FM model, and we also proposed a novel model, named FM-GI, to improve FM model. Experiment result on real dataset shows that this model could achieve higher accuracy , lower computational complexity in rating prediction problem. 2) The final mission of recommender system is to provide a ranked item lists to user. We propose a learning to rank based hybrid model, called LTR-HCF to ensemble social network information and traditional collaborative filtering model. This model has better generality and scalability than the state-of-art social recommendation algorithms. Experiment result shows that LTR-HCF can make more accurate ranking prediction in Top-N recommendation, and also solve cold start user problem partially.
修改评论