Probabilistic Personalized Recommendation Models for Heterogeneous Social Data.

摘要

Content recommendation has risen to a new dimension with the advent of platforms like Twitter, Facebook, FriendFeed, Dailybooth, and Instagram. Although this uproar of data has provided us with a goldmine of real-world information, the problem of information overload has become a major barrier in developing predictive models. Therefore, the objective of this Thesis is to propose various recommendation, prediction and information retrieval models that are capable of leveraging such vast heterogeneous content. More specifically, this Thesis focuses on proposing models based on probabilistic generative frameworks for the following tasks: (a) recommending backers and projects in Kickstarter crowdfunding domain and (b) point of interest recommendation in Foursquare. Through comprehensive set of experiments over a variety of datasets, we show that our models are capable of providing practically useful results for recommendation and information retrieval tasks.