摘要
【目的】利用领域专家信任和相似度相结合的优势,弥补传统协同过滤推荐算法在推荐准确度以及挖掘长尾商品方面存在的不足。【方法】选取Movie Lens中稀疏度为0.9605的数据集,由评分记录较多的1 102个用户对2 920部电影的评分记录构成,利用分阶段实验法求得最优专家用户数量及推荐权重系数?值,并结合对比分析法对算法的性能进行评测。【结果】实验结果表明,本算法的推荐结果准确率和覆盖率均受到专家用户数量的影响,且当推荐权重系数为0.6时推荐准确度明显优于传统算法,同时专家用户比例由2%上升至20%时,覆盖率上升了0.21,说明算法在一定程度上显著提高了推荐系统挖掘长尾商品的能力。【局限】未考虑到不同领域类别之间可能存在的相关性。【结论】该算法能够有效地克服数据稀疏性和冷启动问题,显著提高推荐系统的推荐质量和准确度。
[Objective] This paper tries to improve the performance of traditional collaborative filtering and recommendation algorithm. [Methods] We used the MovieLens dataset to evaluate the proposed algorithm. First, chose datasets with sparse degree of 0.9605, which included scoring records of 1,102 users for 2,920 movies. Second, identified the optimal number of expert users and recommended weight coefficient alpha value with series of experiments. Finally, evaluated the algorithm＇s pertbrmance with comparative method. [Results] The precision of the algorithm were influenced by the expert users. When the recommended weight coefficient value was 0.6, the precision of the new algorithm was better than the traditional ones. Once the propotion of expert users increased from 2% to 20%, the coverage value increased by 0.21. Thus, the new algorithm could analyze the long tail goods more effectively. [Limitations] We did not take into account the possible correlation among different categories. [Conclusions] The proposed algorithm could effectively solve the data sparsity and cold start issues, which significantly improve the performance of the recommendation system.
出处
《现代图书情报技术》
CSSCI
2016年第7期101-109,共9页
New Technology of Library and Information Service
基金
国家社会科学基金项目"数字图书馆标签系统的语义挖掘研究"(项目编号:12CTQ003)的研究成果之一
关键词
个性化推荐
协同过滤
领域专家
相似度
Personalized recommendation Collaborative filtering Domain-Expert Similarity
