UC Santa Barbara

The diffusion of ideas, innovations and behaviors is a complex sociological phenomena for which various computational models have been proposed by economists, algorithmists and ML researchers. However, it remains insufficiently understood due to lack of propagation data on networks and the complex interplay between dynamic network structure and node-level features. Moreover, real-world networks exhibit skewed degree distributions and propagation size distributions, thereby necessitating learning models that account for imbalanced distributions and data scarcity. The goal of this thesis is two-pronged: (1) How do coupled representation and classifier learning models cope with imbalanced data distributions, particularly with minority classes, and how can their weaknesses in limited data settings be mitigated?, (2) How can dynamics of propagation on real-world networks, themselves admitting a natural power law degree distribution, be understood using the machinery of gradient-based optimization and high-dimensional feature representations? The two goals are naturally coupled; distributional analysis is a powerful tool for network (propagation) analysis, and network propagation modeling suffers from a lack of ground truth data, which necessitates a combination of data-driven and synthetic models rooted in sociological insights. Therefore, the major theme of this thesis is models of propagation in networks, and the minor is the limited data learning setting.