Unsupervised Learning in Networks, Sequences and Beyond

Abstract

Unsupervised learning has gained tremendous interest in the past decade in various research communities, by virtue of its capability of exploiting unlabeled data and discovering patterns. The output of classical unsupervised models such as clustering and summarization can serve for the purposes of interpreting as well as preprocessing data for downstream tasks. Unsupervised models also present great potential in representing unstructured data and generalizing to unobserved data. In this thesis, I show the power of unsupervised learning in different data formats - networks and sequences, where models are formed in various ways. For networks, I introduce CondeNSe for summarizing large graphs in an unsupervised way, as well as its relations to graph clustering. I extend the clustering problem to sequences and propose to solve a coupled clustering problem on graphs and sequences. In the sequence modeling domain, I discuss my study on predicting sequences in the context of game theory, which falls in the domain of adversarial learning. Finally, I switch gears to structured data and demonstrate the power of unsupervised models on unstructured data in representing and extracting information.