Improving Rare Case Prediction with Replication Technique

The ability to predict correctly rarely occurring cases is important to the success of applying data mining method to many real life applications. In the context of data mining, rare cases refer to labeled data instances that are infrequently occurred in the database. Discovering infrequent patterns are of interest in some specific domains such as genetic mutant identification, fraud credit card detection, network intruder prevention. But most learning algorithms are biased toward the majority cases such that the minority cases are considered as noise and thus they are ignored during the model induction steps. This ignorance causes the learning algorithm to generate a model that cannot classify or predict a minority case. We thus study the replication technique based on the over-sampling method to solve this problem. However, a straightforward application of oversampling method may lead to the over-fitting problem in such a way that the generated model is too specific to the manipulated data. We thus apply the cluster-based technique to selectively filter a training dataset. The experimental results on primary tumor, arrhythmia and communities-and-crime datasets show significant improvement on predicting accuracy, specificity, and sensitivity of the induced models. But the results on multiple features correlation dataset show non-significant improvement; this case requires further investigation.