Exploiting Context in Feature Selection

Most widely-used feature selection methods assume that features are either relevant in the whole instance space or irrelevant throughout. However, it can often be the case that features are relevant only in the context of other features (e.g., feature Y is relevant if feature X = 1, but irrelevant if X = 0). RC is a new feature selection algorithm that takes this into account, by potentially selecting a di erent set of relevant features for each training instance. When applied to an instance-based learner, it produces higher accuracies than forward and backward sequential selection on a large number of domains, and its advantage increases with increasing context dependency.