FLOWER: Fusing global and local associations towards personalized social recommendation

•We found potential social relations are related with common connections.•We demonstrated the necessity of fusing global and local associations.•We developed a novel method FLOWER to integrate global and local associations.•Methods based on FLOWER significantly outperform others in social recommendation.•We demonstrated the effectiveness of FLOWER on five social networks.

In big data era people are dependent on a variety of social media to manage their social circles. Many online social networks employ social recommendation as an increasingly important component. Although global and local recommendation methods have achieved remarkable success, current studies seldom consider to play advantages of both in social networks. To demonstrate the effectiveness of incorporating local methods to global ones, we first investigated associations between triangular motifs and existing social relations and found that potential links are related with common relations. Further, we analyzed correlations of all methods and clustered them and found obvious strong correlations among methods with same type, which demonstrated the necessity of taking advantages of both. Consequently, we proposed a novel method FLOWER which resorts to Fisher's combined probability test to systematically calibrate statistical significance of global and local associations. FLOWER utilizes information of social relations in both local and global scopes, which are less correlated with each other, and therefore imply possibilities of different aspects for a candidate link. We demonstrated the effectiveness of FLOWER by considering each possible pairwise combination of six global approaches with two local methods and performing 10-fold cross-validation experiments on five real social network datasets (Facebook, Last.fm, Epinions, HEP-PH and Delicious). Results show that FLOWER-based methods significantly outperform either their global or local components in accuracy and retrieval performance.