Nargis Pervin

Recently, a new paradigm of social network based recommendation approach has emerged wherein structural features from social network turned out to be an effective measure to improve the efficacy of the algorithms. However, these approaches assume a user is impacted by all his social connections and completely ignore their preferential similarity, which is crucial for personalized recommendations. Herein, we address this pivotal issue and propose a two-stage clustering based matrix-factorization algorithm, "Cluster REfinement on Preference Embedded MF (CREPE MF)" using a subgraph of social network that integrates the preferential similarity score. Clustering has been applied first on the user followed by the item based on ratings. The proposed algorithm has been systematically evaluated with state-of-the-art algorithms in terms of prediction accuracy and runtime complexity using real-world Yelp dataset. Gratifyingly, our approach outperforms the state-of-the-art algorithms with up to 12.97% and 29.60% improvements in RMSE and runtime, respectively.

With the advent of Web 2.0, recommender systems have become a viable means to harness relevant information online. In the past decades, extensive research have been conducted in the field of recommendations — model-based collaborative techniques being the most favored ones. Recently, a new paradigm of trust-based recommendation approach has emerged wherein structural features from social network resulted in an improved efficacy of the algorithms. However, majority of these approaches assume that users' ratings are impacted by all his social connections in friendship network and completely ignore their preferential similarity, which is essential for personalized recommendations. Herein, we address this pivotal issue and propose a two-stage clustering based matrix-factorization algorithm, ‘Cluster REfinement on Preference Embedded MF (CREPE MF)’ using a subgraph of social network that integrates preferential similarity score. Also, an immense surge in mobile device usage has been observed in recent times, thereby paving the way for tracking users' locations en-route to physical entity recommendations. As users' locations are geo-spatially co-located, we extend CREPE MF to Geographical CREPE MF (gCREPE MF) by incorporating geo-spatial influence. These two proposed algorithms have been systematically evaluated with state-of-the-art algorithms in terms of prediction accuracy and runtime complexity using two real-world data sets, namely Yelp and Gowalla. Gratifyingly, our approach CREPE MF outperforms the state-of-the-art algorithms; depending on the underlying data sets it achieves an improvement of 6.50% to 17.93% in accuracy and 11.67% to 74.23% in runtime. Extended model gCREPE MF further achieves 18.06% to 83.44% reduction in runtime without compromising on accuracy.