Classiication of Incomplete Data Using the Fuzzy Artmap Neural Network

The fuzzy ARTMAP neural network is used to classify data that is incomplete in one or more ways. These include a limited number of training cases, missing components, missing class labels, and missing classes. Modiications for dealing with such incomplete data are introduced, and performance is assessed on an emitter identiication task using a database of radar pulses. 1 A taxonomy of data incompleteness Data presented to a classiier, during either the training or testing phases, may be incomplete in one or more ways: 1. Limited number of training cases: It is of interest to know how the performance of the classiier declines as the amount of training data is decreased, so that, e.g., more training data may be gathered, if necessary, before the classiier is put to use. 2. Missing components of the input patterns: For example, the information in the diierent components of the input patterns may come from diierent sensors, one or more of which may be temporarily unavailable. 3. Missing class labels during training: Some of the training data may have missing class labels. This is referred to as \semi-supervised learning" (Demiriz et al., 1999) or \partially supervised clustering" (Bensaid et al., 1996). (\Missing class labels during testing" is, of course, just the usual situation.) 4. Missing classes: Some classes that were not present in the training set may be encountered during testing. When the classiier encounters a pattern belonging to such an unfamiliar class, it should \\ag" the pattern as belonging to an unfamiliar class, rather than make a meaningless guess as to its identity. This may be implemented by using \familiarity discrimination" (Carpenter et al., 1997). (a) Pure familiarity discrimination. As is common practice when evaluating the performance of a classiier, the classiier does not learn during the testing phase. Test patterns which are agged as unfamiliar are not processed further. In addition to high accuracy of classiication of familiar patterns, the quality of the classiier is measured by a high \hit rate"|fraction of familiar-class test patterns correctly declared to belong to classes familiar during testing and classiied (correctly or not)|and low \false alarm rate"|fraction of unfamiliar-class test patterns incorrectly declared familiar by the classiier. (b) Learning of unfamiliar classes (LUC). The classiier continues to learn during testing. When an unfamiliar class is agged, the classiier deenes a new class, and the criteria for familiarity discrimination are adjusted as necessary. Subsequent test patterns may …