Date of Award

5-2018

Degree Type

Dissertation

Committee Chair

Mohammad Al Hasan

Committee Co-Chair

Jennifer Neville

Committee Member 1

David Gleich

Committee Member 2

Xia Ning

Committee Member 3

Alex Pothen

Abstract

In classical machine learning, hand-designed features are used for learning a mapping from raw data. However, human involvement in feature design makes the process expensive. Representation learning aims to learn abstract features directly from data without direct human involvement. Raw data can be of various forms. Network is one form of data that encodes relational structure in many real-world domains. Therefore, learning abstract features for network units is an important task. In this dissertation, we propose models for incorporating temporal information given as a collection of networks from subsequent time-stamps. The primary objective of our models is to learn a better abstract feature representation of nodes and edges in an evolving network. We show that the temporal information in the abstract feature improves the performance of link prediction task substantially. Besides applying to the network data, we also employ our models to incorporate extra-sentential information in the text domain for learning better representation of sentences. We build a context network of sentences to capture extra-sentential information. This information in abstract feature representation of sentences improves various text-mining tasks substantially over a set of baseline methods. A problem with the abstract features that we learn is that they lack interpretability. In real-life applications on network data, for some tasks, it is crucial to learn interpretable features in the form of graphical structures. For this we need to mine important graphical structures along with their frequency statistics from the input dataset. However, exact algorithms for these tasks are computationally expensive, so scalable algorithms are of urgent need. To overcome this challenge, we provide efficient sampling algorithms for mining higher-order structures from network(s). We show that our sampling-based algorithms are scalable. They are also superior to a set of baseline algorithms in terms of retrieving important graphical sub-structures, and collecting their frequency statistics. Finally, we show that we can use these frequent subgraph statistics and structures as features in various real-life applications. We show one application in biology and another in security. In both cases, we show that the structures and their statistics significantly improve the performance of knowledge discovery tasks in these domains.

Download

Share

COinS