基于码本聚类和因子分解机的多指标推荐算法

传统的协同过滤推荐算法普遍存在数据稀疏问题,且仅利用单一综合评分来计算用户相似度,无法找到在多个指标上偏好相似的用户,因而影响推荐的准确度。多指标评分推荐算法力图寻找在多个指标上偏好相似的用户,但是其评价成本高,导致数据稀疏性问题更加严重。为了找到与目标用户在多个指标上偏好相似的用户,提出基于码本聚类的思想来获取用户在各指标上的评分风格信息,然后基于评分风格信息将用户和项目在各指标上进行双向聚类,最后利用因子分解机模型(Factorization Machines,FMs)基于同一簇内的用户、项目、多指标评分信息、评分风格信息进行推荐。实验结果表明,与传统的协同过滤算法和其他多指标推荐方法相比,基于多指标评分信息的因子分解机推荐算法能够在一定程度上缓解数据稀疏问题,提高推荐的准确度。

Multi-criteria Recommendation Algorithm Based on Codebook-clustering and Factorization Machines

The sparsity of user-item ratings is a common problem and the users who share similar preferences on multi-criteria cannot be found by only making use of a single overall rating to calculate the similarity of users in traditional collaborative filtering algorithm,which would affect the accuracy of recommendation.Multi-criteria recommendation algorithm tries to find users who share similar preferences on multi-criteria,but the problem of data sparsity become even worse owing to the high cost of rating.Aim at these problems,we proposed an algorithm which first obtains the information of rating style of users based on the idea of codebook-clustering,and then conducts co-clustering for users and items on each criteria.Finally,this algorithm makes recommendations by factorization machines(FMs) based on users,items,multi-criteria ratings and rating style.The experimental result shows that multi-criteria recommendation algorithm based on codebook-clustering and FMs is able to solve the problem of data sparsity to some extent,thus improving the accuracy of recommendation.