协同过滤推荐系统是应用最广泛的推荐算法之一,但是其面临严重的稀疏性问题和扩展性问题.针对稀疏的评分矩阵难以准确计算相似度的问题,从推荐算法的流程出发,分离候选集生成和评分预测.针对候选集中存在大量弱或不相关的项目和用户感兴趣比例较低的问题,引入关联度,使用关联矩阵生成候选集;评分预测阶段分析相似度对推荐效果的影响,总结现有相似度的不足,提出一种细粒度划分的"S"型相似度来表述理想增长曲线,并在算法流程中融合候选集生成和评分预测.实验结果表明,减小候选集规模为原来的1/3,避免了评分时对无效项目的计算,算法层面上提高了可扩展性,改进的"S"型相似度在推荐准确率上较之前提高了4％,缓解了稀疏性对推荐效果的影响.%Collaborative filtering recommendation system is one of the most widely used recommendation algorithms, but it faces serious sparseness and scalability. To solve the problem that the sparse scoring matrix is difficult to accurately calculate the similarity, the candidate set generation and scoring prediction are separated from the flow of the recommendation algorithm. Aiming at the problem that there are a large number of weak or irrelevant items and low proportion of users' interest in the candidate set, the correlation degree is introduced and the candidate set is generated by using the correlation matrix. In the prediction stage of scoring, analysis of the influence of similarity on recommendation effect and summarization of the existing shortcomings of similarity, we propose a fine-grained "S" type similarity to express the ideal growth curve. The candidate set generation and scoring prediction are fused in the algorithm. The experiment shows that the size of the candidate set is reduced by 1/3, which avoids the calculation of invalid items when scoring, and the scalability is improved at the algorithm level. The improved "S" type similarity is higher than the former in the recommendation accuracy rate. Increased by 4％, eased the influence of sparsity on the recommendation effect.
展开▼
