摘要
协同过滤是一种简单常用的推荐方法,但是当目标数据非常稀疏时,其性能会严重退化,借助与目标数据跨域关联的辅助数据进行跨领域推荐是解决此问题的一种有效途径.已有的跨领域推荐模型大多假设不同领域完全共享一个评分模式,忽略了领域特有评分模式,可能导致推荐性能退化.此外,许多模型基于单一桥梁迁移跨领域信息,正迁移不足.特别是在考虑领域特有被评分模式的前提下,据作者所知目前还没有模型利用项目的共享被评分模式进行跨领域推荐.因此,该文提出一种新的三元桥迁移学习模型,用于跨领域推荐.首先通过评分矩阵的集合分解提取用户的潜在因子和共享评分模式,以及项目的潜在因子和共享被评分模式,在此过程中考虑了领域特有模式,并对潜在因子施加相似性约束;然后利用潜在因子中的聚类信息构造邻接图;最后通过用户端和项目端的基于共享模式、潜在因子和邻接图的三元桥迁移学习联合预测缺失评分.在三个公开的真实数据集上进行的大量实验表明,该模型的推荐精度优于一些目前最先进的推荐模型.
Collaborative filtering is a recommendation method to predict the missing ratings from target users to target items based on the historical rating data of users similar to target users or items similar to target items. Collaborative filtering is simple and usual, but its performance will degrade seriously when the target data is very sparse. Cross-domain recommendation with the help of auxiliary data associated across domains with target data is an effective way to solve this problem. Most existing cross-domain recommendation models suppose that different domains share a rating pattern completely, which ignores domain-specific rating patterns and may result in degradations of the recommendation performance. In addition, many models transfer cross-domain information based on a single bridge, the positive transfer of which is insufficient. In particular, as far as we know, no model has utilized the shared rated pattern of items for cross-domain recommendation under the premise of considering domain-specific rated patterns. Existing cross- domain recommendation models ignore that the shared rating patterns may have different focuses on users and items, and always only utilize the rating pattern of users. In fact, the rating pattern of users focuses on the interest pattern of users, however, the rated pattern of items focuses on the popularity pattern of items. Therefore, we propose a novel triple-bridge transfer （TRBT） learning model for cross-domain recommendation. Firstly we extract latent factor and shared rating pattern of users as well as latent factor and shared rated pattern of items by collective factorizations on rating matrices, while considering domain-specific patterns and imposing similarity constraints on latent factors; then we construct adjacency graphs utilizing clustering in- formation contained in latent factors; finally we predict the missing ratings jointly by user-side and item-side triple-bridge transfer learning based on shared pattern, latent factors and adjacency graphs. During the triple-bridge transfer, the interactive information between users and items is transferred based on the shared patterns, the feature information of users and items is transferred based on the latent factors, and the adjacency information of users and items is transferred based on the adjacency graphs. We highlight the different focuses of the shared pattern on users and items and utilize the shared rated patterns of items for the first time. Our proposed TRBT model increases the positive transfer and meanwhile reduces the negative transfer. Furthermore, TRBT model not only can transfer useful cross-domain information of three different types, but also can choose transfer bridges and adjust transfer weights of different bridges flexibly according to the specific situation. Although TRBT model may seems a bit complex, the core algorithms of it can be solved by iterations with a small number and converge with a fast speed, so it can be applied to cross-domain recommendation under the environment of massive data in theory. We adopt both mean absolute error （MAE） and root mean square error （RMSE）, which are two popular evaluation metrics for collaborative filtering recommendation models at present, to evaluate the predicting accuracy for missing ratings of different recommendation models. Extensive experiments on three public real world datasets demonstrate that the recommendation accuracy of TRBT model outperforms several state-of-the-art recommendation models.
作者
王俊
李石君
杨莎
金红
余伟
WANG Jun LI Shi-Jun YANG Sha JIN Hong YU Wei(State Key Laboratory of Software Engineering, School of Computer Science, Wuhan University, Wuhan 430072 School of Computer Science and Technology, Hankou University, Wuhan 430212)
出处
《计算机学报》
EI
CSCD
北大核心
2017年第10期2367-2380,共14页
Chinese Journal of Computers
基金
国家自然科学基金项目(61272109)
中央高校基本科研业务费专项资金项目(2042014KF0057)
湖北省自然科学基金项目(2014CFB289)资助~~
关键词
迁移学习
推荐
协同过滤
跨领域
稀疏
矩阵分解
transfer learning
recommendation
collaborative filtering
cross-domain
sparse
matrix factorization
