•   When: Thursday, November 13, 2014 from 10:00 AM to 12:00 PM
•   Speakers: Tanwistha Saha
•   Location: ENGR 4201
•   Export to iCal

Abstract

Graphs, a.k.a. relational networks, have emerged as a powerful mechanism for representing complex structured data captured from the omnipresent social and information networks. The past decade has seen a tremendous growth in web services and applications that allow users to communicate with each other, to collaborate in creating media, discuss and rate purchased entities and items. Relational networks are also found in a variety of domains ranging from gene-regulatory networks, author-citation networks, and inter-country trade networks. The objective of this dissertation is to develop approaches for extracting useful information from these social and information networks.

Specifically, I present classification algorithms for predicting class labels associated with nodes of relational networks. Acquiring class labels for these nodes is an expensive process. As such, I introduce novel active learning based approaches for acquiring class labels and training relational classifiers. Our techniques apply to both single-labeled and multi-labeled networks. Our experimental evaluation on real-world data shows statistically significant results on both single- and multi-labeled networks in comparison to state-of-the-art approaches.

Relational networks play an important role also for recommender systems. They capture the inter-connections between various users and items. Traditional recommender systems identify and recommend interesting items to a given user based on the user's past rating activity. These systems improve their recommendations by identifying user preferences and item related information from external sources, like reviews written by users, or concept tags about items shared by users. In this dissertation, I also seek to improve recommender systems by integrating user preferences as side information within standard neighborhood-based and latent factor based methods. We assume that a user's choice of tags provides additional information about the user's personal preference and additional features about the item. Since querying users to provide tags imposes an additional burden on the user, we demonstrate the use of relational classification algorithms for predicting tags for both items and users. Our experimental results on several real-world data show the advantages of using tag-based information within recommender systems.