Top-K Item Recommendations Using Social Media Networks - Using Twitter Profiles as a Source for Recommending Movies

The advent of internet has served as an offspring for the significant growth of online services and businesses such as e-commerce, entertainment, or social media. A common element among these industries is the process of tailoring the offered services or products towards their users' interests and preferences, also known as personalization. Related to this is the cold start problem, wherein systems may not have sufficient data on new users or customers in order to provide reasonable, personalized recommendations. In an attempt to overcome said challenge, this thesis investigates the use of user data available from social media - in this case public Twitter profiles. A two-step recommender system is proposed and implemented, using the aforementioned data as input and basis for its predictions. The first step of our approach involves classifying and filtering Tweets based on their expressed sentiment, using Artificial Neural Networks to achieve state-of-the-art classification performance. Following this, we experiment with various combinations of recommender algorithms in order to match a specific user's preferences with the aspects of any movie. The experiments examine the impact of numerous variables, including preprocessing techniques, feature extraction, similarity measures, word embeddings, entity matching and social circles with regards to the systems predictions. The results produced from the recommender system for any given user is a ranked list of movie titles with corresponding similarity scores. We evaluate the proposed system on a larger set of Tweets annotated with sentiments, a set of Twitter user profiles, as well as a set of movies with associated data from IMDb. A prediction accuracy of 65% in the most successful case was reached, in which user-based collaborative filtering was utilized. Overall, the results of our experiments indicate that the use of social media profiles has merit in the task of movie recommendations, and may have applications in other domains as well