Active Learning for Multi-Class Logistic Regression

Which of the many proposed methods for active learning can we expect to yield good performance in learning logistic regression classifiers? In this article, we evaluate different approaches to determine suitable practices. Among our contributions, we test several explicit objective functions for active learning: an empirical consideration lacking in the literature until this point. We develop a theoretical framework for applying different loss functions motivated by work in optimal experimental design. Empirical investigations demonstrate the benefits of our variance reduction method which gives attractive classification accuracy and matches or beats random performance in all evaluations. Of the alternative heuristic approaches, we identify a method called margin sampling as giving promising performance with little computational overhead.