| In the real world, concept drift happens in various scenarios including medical treatment recommendation, where the relation between features and the target class changes over time in unforeseen ways. Nearest neighbors(k-NN) is a simple non-parametric classification model, yet it is effective
in various situations, even for training data with an unknown distribution. While categorical attributes limit the application of many learning algorithms, k-NN classifier still behaves well in such scenarios given a reasonable similarity measure between categorical attributes. To alleviate deterioration in prediction accuracy introduced by concept drift, sliding windows and temporal weights are often used, simply to eliminate/dilute the effect of outdated samples on the prediction, leading to a less confident (based on fewer samples) prediction and a waste of undiscovered information contained in past samples. With the knowledge of how concepts change, outdated samples can be adapted for up-to-date prediction, which improves the confidence
of prediction, especially for medical data sets of which the scale is relatively small. In this paper we present an adaptive k-NN classifier which can detect the occurrence of target concept drift and update past samples according to the knowledge of the drift for better prediction, and evaluate its performance over both simulated and real categorical medical data sets. The experiment results show our classifier achieves better performance under concept drift. |
