异质社交网络中协同排序的好友推荐算法

Collaborative Ranking Friend Recommendation Algorithm in Heterogeneous Social Network

随着社交网络的复杂化和异质化,传统推荐系统中协同过滤推荐方法由于推荐效果不佳而不能满足需求.本文通过扩展原有推荐方法中的因子模型提出了基于协同排序的好友推荐算法.相比于协同过滤,本文使用用户之间的偏序关系取代原始打分,以适合不易评分的异质信息网络,并且对于Top-k推荐只需考虑推荐序列,不需要精确预测低序列的评分的特点,避免不必要的计算,提高计算效率.相对于普通的因子模型,本方法在好友推荐中训练集更易构建,可以简单有效的融合多种有价值的内容相关特征.测试数据表明,基于协同排序的好友推荐与以往的矩阵分解方法相比较,在Digg2009好友关注关系数据集上测试,MAP提高了15.6%左右.

The social network is more and more complex and heterogeneous, traditional friend recommendation algorithm cannot cope the new situation for its ineffectiveness. This paper proposes a collaborative ranking based friend recommendation method by expending the traditional factor model. Comparing to the collaborative filtering friend recommendation methods, we replace the original scoring method with the partial order relation between the users to satisfy the needs of heterogeneous social network which is suitable for some situations that are bard to convert rates and it＇s of no need to precisely calculate the rates of low users in Top-k recommendation so that it is helpful to enhance the efficiency. According to the experiment results, our method is easy for friend recommendation to build training data and gets better results in learning user＇s interest than factorization model. The method can also mixes valuable content related feature easily. After testing in dataset of Digg2009, Collaborative ranking get 15.6% higher than Matrix factorization in MAP.