考虑用户间消极相似性的排序推荐算法

Ranking-based recommendation algorithm considering negative similarity between users

由于用户评分标准存在差异,基于打分的协同过滤推荐算法在近邻选择过程中存在误差。针对以上问题,提出考虑用户间消极相似性的排序推荐算法(NS-TauRank),该算法不经过对拟推荐项目的预测评分过程。定义DP函数表示项目对相关属性,充分利用用户间的消极相似性,即相似性为负的用户之间的爱好相反,改进目标用户的近邻选择过程,采用舒尔茨方法进行偏好融合,优化目标用户拟推荐项目的排序。在Eachmovie和movielens数据集上对改进算法进行验证,以NDCG作为评价函数,验证结果表明,该算法在两个数据集上的NDCG@1-2值较对比算法有4%-7%的提高,产生了更可靠的拟推荐序列。

When the rating standards between users are various, neighbors, selection using the traditional collaborative filtering algorithms is quite different. To solve the problem, a ranking-based recommendation algorithm considering negative similarity between users （NS-TauRank） was proposed to directly address the item ranking problem without going through the inter-medi- tate step of rating prediction. A DP function was defined to weight the attribute of items. Negative similarity was proposed to make full use of the impact from the neighbors with opposite interest to the target user. Schulze method was used in order-based preference information aggregation. Experimental results on two real movie rating datasets show that the value of NDCG@1-2 is 4%- 7 % higher than normal algorithms , which means the proposed approach outperforms traditional collaborative filtering algorithm.