Abstract

The growing access to large amounts of structured observations allows for more opportunistic uses of this data. An example of this, is the prediction of an event's class membership based on a database of observations. When these predictions are supported by a highlevel representation, we refer to these as knowledge based on-line classification tasks. Two common types of algorithms from machine learning research that may be applied to on-line classification tasks make use of either lazy instance-based (k-NN,IB1) or eager model-based (C4.5,CN2) approaches. Neither approach, however, appears to provide a complete solution for these tasks. This thesis proposes a lazy model-based algorithm, named DBPredictor, that is suited to knowledge based on-line classification tasks. The algorithm uses a greedy top-down search to locate a probabilistic IF-THEN rule that will classify the given event. Empirical investigation validates this match. DBPredictor is shown to be as accurate as IB1 and C4.5 a...

Citations