A Fusion Approach for Collaborative Filtering
Soumita Das, Sambo Dutta · 2019
Collaborative Filtering (CF) is one of the most important memory based techniques used in recommender systems. CF exploits interests of all users in the database similar to that of the target user to recommend him/her items. However as the number of users and items grow in the database, it becomes difficult for traditional CF based recommender systems to search for similar users in the entire user space. This gives rise to accuracy and scalability issues and the CF algorithm fails to give quality real time results. In this paper we propose a fusion approach of leveraging the geographical location of the users and their social network graph to increase the accuracy of CF algorithm. We first cluster users based on their geographic location with the intuition that users within a same region usually have similar preferences and habits. Secondly, we use the social network graph to find friends of the target user as it is likely that a person will have more inclination towards his/her friend's recommendation than a stranger's. Lastly, we form a fusion of the target user's social network friends and other users who fall in the cluster of the target user and execute the CF algorithm. A Distance Threshold Filter is applied to generate the final recommendation list for each user. Results using evaluation metrics like MAE and RMSE confirms that the fusion approach provides better quality recommendations than other state-of-art collaborative filtering algorithms.