Machine Learning and Friends Lunch
Non-Bayesian Networks
Ron Bekkerman
UMass
Abstract
Generative models constructed for density estimation problems are often enormous in size and heavily biased. Still, they remain extremely popular for unsupervised learning within the graphical model framework. Other types of graphical models are emerging. In this talk I will discuss ongoing work on defining Non-Bayesian Networks - undirected graphical models that go beyond the scope of probabilistic learning. I will propose the main principle of non-Bayesian inference and will then present an instance of non-Bayesian networks called a Combinatorial Markov Random Field (Comraf), which is a compact, data-driven, generic model that can be applied to information retrieval, data mining, computer vision, collaborative filtering and other domains. I will show that on the task of document clustering the Comraf model significantly outperforms other models, both generative (such as LDA) and information-theoretic. I will also show that Comrafs are straightforwardly applicable to semi-supervised learning and transfer learning. Finally, I will briefly discuss other Comraf applications such as topic detection, image clustering and multi-way ranking.
Joint work with Mehran Sahami.