Recommendation system for Facebook public events based on probabilistic classification and re-ranking

The evolution of Facebook social network with event feature helps improve the quality of interaction among users in real life. Different from other objects such as movies and books, recommendation problem for events is inherently cold-start and affected by the natures of each social network platform (Facebook, Meetup, etc). In this paper, we discuss the character of Facebook social network and how the event recommendation problem in this case is different to others. We argue that rather than just trying to place as many as possible the most suitable events (often based on similarity measurement) on top recommendations, it's better to remove the unsuitable ones and reorder the remaining in a way that improves the user experience. From that, we propose a new method for event recommendation divided into two consecutive stages - classification and re-ranking. For the first phase, we use a blending model of probabilistic classifiers to predict positive and negative probabilities for each user-event pair, then evaluate on those results to eliminate all bad cases before passing the rest to the next phase. We also propose a new optimization procedure in this evaluation process. For the second phase, we treat the positive probability as a measurement of similarity and make some comparisons across several reranking techniques to choose the best based on the objective of improving quality of recommended lists. Experimental results on crawled Facebook public events show the effectiveness of the proposed method.