有向相似性对协同过滤推荐系统的影响研究

摘要

In this paper, to study the effect of user similarities to CF recommendation algorithms, we argue that the similarities which should be taken into account are those come from the neighbor users to the target user. Based on the above idea, we present a modified CF algorithm. The numerical results on a benchmark dataset, MovieLens, show that by using the direction from neighbor users to the target user, the performance of this algorithm, including accuracy and diversity, can be improved greatly. More importantly, we find that when enhancing the higher similarity users' recommendation power, the accuracy can reach 0. 086 4, which is further improved by 17. 94%. When the recommendation length equals to 10, the diversity reaches 0. 892 9 and be further improved by 20. 9%. Our work indicates that the direction of user similarity is an important factor of the CF algorithm.%为研究用户的相似性对协同过滤个性化推荐算法的影响,认为用户的有向相似性应该由邻居用户指向目标用户,而非由目标用户指向邻居用户.基于该思想,提出了一类改进的协同过滤算法.通过对Movielens数据集的实验分析,结果发现改变用户相似性的方向可大幅提高推荐结果的准确度和推荐列表的多样性.进一步,强化相似度高的用户的推荐强度可大幅提高推荐效果,算法的准确性可提高17.94％,达到0.086 4,当推荐列表的长度为10时,推荐列表的多样性可达到0.892 9,提高20.9％.该工作表明用户相似性的方向是否合理对推荐算法具有非常大的影响.