Title: “Multi-label Classification via Binary Markov Networks”
Advisors: Associate Professor Ji Zhu, Professor Vijay Nair
Committee Members: Associate Professor Liza Levina, Professor Peter X.K. Song (Biostatistics)
Abstract: Multi-label classification refers to the scenario in classification that each instance is associated with a subset of labels rather than one. The labels are not mutually exclusive and often correlated. In this project, we first propose to transform multi-label classification into a multivariate binary regression problem. Then we introduce an Ising model with covariates to explicitly model the conditional distribution of the class labels given the covariates. Pseudo likelihood is adopted to develop a computationally efficient estimation procedure. We also investigate the choice of evaluation measures in connection to different prediction rules, which is further illustrated by numerical studies. The proposed method is applied to a popular Yeast dataset and shows promising result.