University of Alberta DECISION TREE INSTABILITY AND ACTIVE LEARNING

The instability of learning algorithms is an important topic that is often overlooked in the field of machine learning. An unstable learner is one that may yield dramatically different classifiers from training sets that differ just slightly. Decision tree learning algorithms produce accurate models that can be interpreted by a domain expert with relative ease. However, these algorithms are known to be highly unstable, which undermines the common objective of extracting knowledge from the tree structures they create. This thesis examines the instability of the C4.5 decision tree learner in two contexts: passive learning and active learning. An alternative splitting criterion is tested with C4.5, in order to determine whether it improves the stability of the algorithm. Several active learning methods that use C4.5 as a component learner are compared with respect to the stability and accuracy of the decision trees they produce.