摘要
基于共同评分与项目全集的相似度未甄别近邻的推荐影响力,导致推荐质量低,可扩展性差。为此,提出了一种基于推荐影响度的并行协同过滤算法。该算法通过非共同评分项目、共同评分项类以及用户访问次数来计算用户推荐新颖度与兴趣重合度以度量用户推荐能力,并融入相似性计算来抑制相似度高但推荐力不强的用户,避免在项目全集上计算相似度,从而提高推荐质量;通过MapReduce并行化,使其具备良好的实时性和可扩展性。实验结果表明,该算法在海量数据集上的推荐质量更高,可扩展性更强。
The similarity based on common scores and full item sets has failed to identify the nearest neighbor recom- mendation influence,which brings about lower recommend quality and poor scalability. Through non-common rating items,common score item categories and user visited times, this paper proposed a parallel collaborative filtering algo- rithm based on user recommendation influence. It computes the user recommended novelty degree and interest coinci- dence to measure user recommendation influence ability. By adding it to calculate similarity, the algorithm can effectively restrain the highly recommended users with high similarity, avoid similarity computation on full item sets and improve the quality of recommendation. Further more, by using MapReduce parallelization, this algorithm has good real-time per- formance and scalability. The experimental results show that the parallel algorithm is of higher recommendation quality and better scalahility on big data.
出处
《计算机科学》
CSCD
北大核心
2017年第9期250-255,271,共7页
Computer Science
基金
河北省高等学校科学技术研究重点项目(ZD2014061)
青年基金项目(QN2016108)资助
关键词
推荐影响度
推荐新颖度
兴趣重合度
MapReduce并行化
Recommendation influence degree,Recommendation novelty degree,Interest coincidence degree,MapReduce paralleliation