Multiple social network integration framework for recommendation across system domain

Abstract

A recommender system is a special software that recommends items to a user based on the user’s history. A recommender system comprises users, items and a rating matrix. Rating matrix stores the interactions between users and items. The system faces a variety of problems among which three are the main concerns of this research. These problems are cold start, sparsity, and diversity. Majority of the research use a conventional framework for solving these problems. In a conventional recommender system, user profiles are generated from a single feedback source, whereas, Cross Domain Recommender Systems (CDRS) research relies on more than one source. Recently researchers have started using “Social Network Integration Framework”, that integrates social network as an additional feedback source. Although the existing framework alleviates recommendation problems better than the conventional framework, it still faces limitations. Existing framework is designed only for a single source domain and requires the same user participation in both the source and the target domain. Existing techniques are also designed to integrate knowledge from one social network only. To integrate multiple sources, this research developed a “Multiple Social Network Integration Framework”, that consists of two models and three techniques. Firstly, the Knowledge Generation Model generates interaction matrices from “n” number of source domains. Secondly, the Knowledge Linkage Model links the source domains to the target domain. The outputs of the models are inputs of the techniques. Then multiple techniques were developed to address cold start, sparsity and diversity problem using multiple source networks. Three techniques addressed the cold start problem. These techniques are Multiple Social Network integration with Equal Weights Participation (MSN-EWP), Multiple Social Network integration with Local Adjusted Weights Participation (MSNLAWP) and Multiple Social Network integration with Target Adjusted Weights Participation (MSN-TAWP). Experimental results showed that MSN-TAWP performed best by producing 47% precision improvement over popularity ranking as the baseline technique. For the sparsity problem, Multiple Social Network integration for K Nearest Neighbor identification (MSN-KNN) technique performed at least 30% better in accuracy while decreasing the error rate by 20%. Diversity problem was addressed by two combinations of the cold start and sparsity techniques. These combinations, EWP + MSN-KNN, TAWP + MSN-KNN and TAWP + MSN-KNN outperformed the rest of the diversity combinations by 56% gain in diversity with a precision loss of 1%. In conclusion, the techniques designed for multiple sources outperformed existing techniques for addressing cold start, sparsity and diversity problem. Finally, an extension of multiple social network integration framework for content-based and hybrid recommendation techniques should be considered future work.