Information overload in structured data

Pinar Yanardag Delul, Purdue University

Abstract

Information overload refers to the difficulty of making decisions caused by too much information. In this dissertation, we address information overload problem in two separate structured domains, namely, graphs and text. Graph kernels have been proposed as an efficient and theoretically sound approach to compute graph similarity. They decompose graphs into certain sub-structures, such as subtrees, or subgraphs. However, existing graph kernels suffer from a few drawbacks. First, the dimension of the feature space associated with the kernel often grows exponentially as the complexity of sub-structures increase. One immediate consequence of this behavior is that small, non-informative, sub-structures occur more frequently and cause information overload. Second, as the number of features increase, we encounter sparsity: only a few informative sub-structures will co-occur in multiple graphs. In the first part of this dissertation, we propose to tackle the above problems by exploiting the dependency relationship among sub-structures. First, we propose a novel framework that learns the latent representations of sub-structures by leveraging recent advancements in deep learning. Second, we propose a general smoothing framework that takes structural similarity into account, inspired by state-of-the-art smoothing techniques used in natural language processing. Both the proposed frameworks are applicable to popular graph kernel families, and achieve significant performance improvements over state-of-the-art graph kernels. In the second part of this dissertation, we tackle information overload in text. We first focus on a popular social news aggregation website, Reddit, and design a submodular recommender system that tailors a personalized frontpage for individual users. Second, we propose a novel submodular framework to summarize videos, where both transcript and comments are available. Third, we demonstrate how to apply filtering techniques to select a small subset of informative features from virtual machine logs in order to predict resource usage.

Degree

Ph.D.

Advisors

Vishwanathan, Purdue University.

Subject Area

Information Technology|Computer science

COinS