Finding the most informational friends in a Social Network based Recommender System

Information present in social networks if incorporated in recommender systems, can lead to highly accurate and meaningful recommendations. Further, if social influence of only the most informational friends is taken into account, accuracy is further improved. The most informational friends are those friends who have had similar preferences to the target user in the past. These friends provide the most relevant information to compute accurate recommendations. Social Network-based Recommender System (SNRS) [6] is a probabilistic model that takes into account user preferences, item's general acceptance and the friends' influence to provide recommendations. It overcomes the limitations of data sparseness and cold-start problem which are characteristic of the traditional algorithms such as collaborative filtering. This paper proposes six different techniques for finding the most informational friends in the SNRS framework with an aim to improve the overall accuracy. Statistical tools like mean absolute error, correlation, entropy, clustering and graph partitioning are used in the probabilistic framework to grade the friends on the basis of information contributed. The optimum number of informational friends is determined both globally as well as dynamically on a per user basis. Extensive experimentation on an existing social network for restaurant recommendations establishes the usefulness of our approach.