基于社会信任正则化的排名推荐算法

随着在线商品交易额逐年增大和社交网络不断深入发展，推荐系统已成为解决信息过载的重要工具之一。当评分矩阵数据稀疏性较大时推荐精度就会显著下降，特别是用户冷启动的时候这个问题更加明显。因此，本文提出一种新的基于隐式反馈信息的社会化排序推荐算法。该算法首先利用矩阵分解方法计算不同项目间的用户偏好。其次将用户偏好信息融入Bayesian Personalized Ranking (BPR)算法中。然后挖掘用户之间相似关系以及信任用户直接和间接关系，并量化它们之间的信任关系，从而研究不同项目之间用户偏好差异。最后将以上这些信任关系和BPR算法进行融合，进而构建出社会化排序推荐模型。为了验证所提出的社会化排序推荐算法，在DouBan数据集和FilmTrust数据集上，进行该算法的有效性验证。主要通过Precision、MAP和NGCD这三种排序评估指标分别在全数据集和用户冷启动中验证本文所提算法与SBPR、TBPR、BPRMF和MostPopular等算法之间排序推荐的优劣性。实验结果证明本文所提算法明显优于其他排序推荐算法，并可以获得更好的推荐准确率。可见该算法可以有效改善由于数据稀疏性和用户冷启动所带来推荐效果差的问题。

Ranking Recommendation Algorithm Based on Social Trust Regularization

With the increase of online commodity transaction volume and the further development of social network, recommendation system has become one of the important tools to solve information overload. However, when the data sparsity of the score matrix is relatively large, the recommendation accuracy will decrease significantly, especially the user starts cold. A novel pairwise learning to socialized ranking recommendation algorithm based on implicit feedback information is proposed in this article. The algorithm first calculates user preferences among different items using a matrix factorization method. Secondly, user preference information is incorporated into the Bayesian Personalized Ranking (BPR) algorithm. Similar relationships between users as well as direct and indirect relationships between trusting users are then mined and quantified in order to study differences in user preferences across projects. Finally, these trust relationships are fused with the BPR algorithm to build a social ranking recommendation model. To validate the proposed social ranking recommendation algorithm, the validity of the algorithm is verified on the DouBan dataset and FilmTrust dataset. The ranking evaluation metrics of Precision, MAP, and NGCD are used to verify the merits of ranking recommendation between the proposed algorithm in this paper and SBPR, TBPR, BPRMF, and MostPopular algorithms. Ranking recommendation tests are performed on these two specific social datasets for the full dataset and user cold start, respectively. The experimental results demonstrate that the proposed algorithm in this paper significantly outperforms other ranking recommendation algorithms and can achieve better recommendation accuracy.It can be seen that the algorithm can improve the problem of poor recommendation due to data sparsity and user cold start.