基于正相关和负相关最近邻居的协同过滤算法
Collaborative Filtering Algorithm Based on Positive Correlation and Negative Correlation Nearest Neighbors
作者：徐怡(安徽大学 计算机科学与技术学院, 安徽 合肥 230601;计算智能与信号处理教育部重点实验室(安徽大学), 安徽 合肥 230039)；唐一民(安徽大学 计算机科学与技术学院, 安徽 合肥 230601)；王冉(安徽大学 计算机科学与技术学院, 安徽 合肥 230601)
Author：XU Yi(School of Computer Sci. and Technol., Anhui Univ., Hefei 230601, China;Key Lab. of Intelligent Computing & Signal Processing(Anhui Univ.), Ministry of Education, Hefei 230039, China)；TANG Yimin(School of Computer Sci. and Technol., Anhui Univ., Hefei 230601, China)；WANG Ran(School of Computer Sci. and Technol., Anhui Univ., Hefei 230601, China)
收稿日期：2017-09-17 年卷（期）页码：2018,50(5):189-195
期刊名称：工程科学与技术
Journal Name：Advanced Engineering Sciences
关键字：协同过滤;正相关邻居;负相关邻居
Key words：collaborative filtering;positive correlation nearest neighbors;negative correlation nearest neighbors
基金项目：国家自然科学基金资助项目（61402005）；安徽省自然科学基金资助项目（1308085QF114）；安徽省高等学校省级自然科学基金资助项目（KJ2013A015）；安徽大学计算智能与信号处理教育部重点实验室课题项目资助；国家留学基金委员会资助项目
中文摘要
协同过滤算法是应用最广泛和最成功的推荐算法之一。针对传统协同过滤算法在预测评分时仅考虑正相关最近邻居，没有考虑负相关最近邻居对评分预测的影响而导致的预测结果准确性及多样性较低的问题，提出一种基于正相关和负相关最近邻居的协同过滤算法。该算法首先计算用户之间的相似度，再通过用户评分与其平均评分等信息计算出用户之间的变异系数，利用变异系数修正相似度的值，从而缓解因为用户共同项目数不足而导致的相似度计算结果可信度较低的问题。然后分别对与目标用户相似度为正及与目标用户相似度为负的用户进行排序，并利用动态加权参数α及训练得到的阈值β分别选取正相关最近邻居和负相关最近邻居，基于选取的正相关最近邻居和负相关最近邻居分别进行预测评分。最后，将基于正相关最近邻居和负相关最近邻居的预测评分进行加权，作为最终的预测评分。在MovieLens数据集上利用3种评价标准进行对比实验，结果表明本文算法有效地提高了推荐的准确性和多样性。
英文摘要
Collaborative filtering algorithm is one of the most widely and successfully used recommendation algorithm. Traditional collaborative filtering algorithm only considers the positive correlation nearest neighbors while ignore the influence of negative correlation nearest neighbors when predicting item rating, which has caused problems like low accuracy and low diversity. In order to resolve these problems, based on positive correlation and negative correlation neighbors a collaborative filtering algorithm is proposed. Firstly, the algorithm computes the similarities and Coefficient of Variation between users and sorts the neighbors according to the similarities which were modified by the Coefficient of Variation. Secondly, the positive nearest neighbors and the negative nearest neighbors are selected based on dynamic weight valueαand threshold valueβ, then item ratings are predicted based on the positive nearest neighbors and the negative nearest neighbors respectively. Finally, the last prediction result is acquired by combining predicting ratings based on the positive nearest neighbors and the negative nearest neighbors with weight. The experiments are tested on the MovieLens with three metrics, and the result show that this approach is achieved great improvement of the accuracy and diversity of recommendation effectively.
【关闭】
