Link prediction models for recommendation in academic collaboration network of Turkey

Official URL: http://risc01.sabanciuniv.edu/record=b1655731 (Table of Contents)

Recommendation of individuals in a professional network can be thought as a link prediction problem. For this purpose, learning based methods are suitable for both prediction and inference goals. These methods are also easy to implement and they are computationally tractable for large scale datasets. In this thesis, we investigate link prediction models for recommendation of authors and papers in academic collaboration network of Turkey. We examine the dataset with certain visualization tools and basic statistics, in order to form a conceptual background for our models and hypothesis. As a byproduct, we highlight potential drawbacks of existing models and the information resources they use. Taking the conceptual background and potential drawbacks into account, we implement, compare and contrast existing models, and test our hypothesis. We cast all of these models as optimization problems and solve them by using first order optimization methods. We evaluate their results in terms of prediction accuracy and provide interpretation about their meaning as well. While the results that we have obtained are special for the given dataset that we have, techniques that we have used are generic enough to be implemented with other graph datasets.