Institute for Software Research
School of Computer Science, Carnegie Mellon University
Bayesian Mixed-Membership Models
Edoardo Maria Airoldi
Ph.D. Thesis (COS)
A solution to the global/local trade-off is to express complexity through hierarchical mixtures of simple patterns, i.e., motifs, that evolve over time. Complex global behavior emerges from the combination of local interaction patterns and their dynamics. I discuss the extent to which this novel framework incorporates, generalizes, and extends other probabilistic approaches present in the literature, and argue that it provides the foundations of a statistical theory of random graphs.
A major part of the effort is devoted to the analysis of modeling issues related to the four essential aspects listed above, in the context of applications to social and biological networks. I also investigate theoretical and computational issues such as the geometrical intuition of the latent allocation task–an important inference objective shared by the various models encompassed by this framework.