Dynamic feature weighting in nearest neighbor classifiers

One major problem of nearest neighbor (NN) algorithms is inefficiency incurred by irrelevant features. A solution to this problem is to assign weights to features that indicate their salience for classification. Current weighting methods can be divided as global weighting, partial local weighting, and local weighting methods enumerated in increasing order of capability to capture the features' relative salience in classification. However, the existing methods are not sensitive enough to describe the salience of a feature and can be changed given different queries. We suggest that the salience of a feature, in addition to being sensitive to the instance (i.e. varies across instances), should also be sensitive to the variations in the difference of a feature's values between a query and the instances in the instance base. In this paper, we put forward a dynamic feature weighting approach which has more expressive capability, and present a sketch of a classification algorithm based on the notion of dynamic weights.