Abstract:
In this thesis, different frameworks for a recommendation system based on social network analytics is investigated. In these frameworks, three different potential customer identification approaches are examined and corresponding successes are analyzed. In order to exploit the underlying network structure, three networks, restaurant-user, user-user and restaurant-restaurant, are generated. In the first approach, potential users are ranked and selected according to a combination of pagerank values and community scores of both restaurants and users. In the second approach, users are ranked according to the sentiments scores of their comments in conjunction with pagerank of restaurants. In the third approach, node embeddings for the restaurant-user network are computed and used to find the similarities between users and restaurants. Then, based on these similarities, potential users are ranked for a given focal restaurant. With the aim of comparing the successes of these three frameworks, dataset is splitted into three and success rates are calculated based on the percentage of the actual customers recommended by the generated models. Experiments in this research shows that Ranks framework utilizing the community structure together with the network ranking of both users and brands reached up to 50% and on average achieved 9.61% accuracy when the number of potential customers to be recommended is taken as 100. So, frameworks utilizing the underlying network structure can be exploited to improve the prediction capability of recommendation systems that find potential customers for a given company or brand.
