Abstract

We approach the problem of active learning from a Bayesian perspective, working with a probability distribution over the solution space. In addition to the classical active selection of data points, we formulate the construction of minimum decision trees from noisy datasets as an active learning task. Building on the OASIS algorithm, we compare active learning score functions based on the EC2 criterion with uncertainty sampling, two GBS approaches, and random selection by performing experiments on several standard datasets. While constituting a unique approach in its original online setting, according to our findings, the OASIS algorithm does not generally offer performance benefits in a classical offline setting. Furthermore, for active learning from noisy data samples, we introduce a new intrinsic EC2-based stopping criterion and show that in many cases, it outperforms a standard information gain based method paired with the χ2-test on the decision tree learning problem. In particular, our criterion enables the effective construction of small decision trees, and may provide the first effective method to learn a close-to-minimal tree classifier with bounded expected error rates from noisy data samples. i Acknowledgments I want to thank Andreas Krause for his caring supervision, in particular for all the time, effort and energy spent on providing guidance and explanations. Furthermore, I am indebted to Victor Candia for his support and helpful advice, and to Wito Traub for his comments. My parents, Josef and Madeleine Rupprechter, deserve special thanks along with the rest of my family for their ongoing support, without which the writing of this thesis would have been impossible. Last but not least, a big thank you to the two loves of my life, Maya Minwary