TAFFY: incorporating tag information into a diffusion process for personalized recommendations

Abstract

The last few years have witnessed an explosion of information caused by the exponential growth of the Internet and World Wide Web, which confronted us with information overload and brought about an era of big data, appealing for efficient personalized recommender systems to assist the screening of useful information from various sources. As for a recommender system with more than the fundamental object-user rating information, such accessorial information as tags can be exploited and integrated into final ranking lists to improve recommendation performance. However, although existing studies have demonstrated that tags, as the additional yet useful resource, can be designed to improve recommendation performance, most network-based approaches take users, objects and tags as two bipartite graphs, or a tripartite graph, and therefore overlook either the important information among homogeneous nodes in each sub-graph, or the bipartite relations between users, objects or tags. Moreover, recent studies have suggested that the filtration of weak relationships in networks may reasonably enhance recommendation performance of collaborative filtering methods, and it has also been demonstrated that approaches based on the diffusion processes could more effectively capture relationships between objects and users, hence exhibiting higher performance than a typical collaborative filtering method. Based on these understandings, we propose a data fusion approach that integrates historical and tag data towards personalized recommendations. Our method coverts historical and tag data into complex networks, resorts to a diffusion kernel to measure the strength of associations between users and objects, and adopts Fisher's combined probability test to obtain the statistical significance of such associations for personalized recommendations. We validate our approach via 10-fold cross-validation experiments. Results show that our method outperforms existing methods in World Wide Web not only the recommendation accuracy and diversity, but also retrieval performance. We further show the robustness of our method to related parameters.