Learning on extremes - size and imbalance - of data .

摘要

The availability of extremely large datasets has opened avenues for application of distributed and/or parallel learning. Our approach to learning from such large datasets is to utilize all the training data by learning classifiers on manageable subsets of data. Using a voting mechanism, the predictions of the individual classifiers can be combined. Our experiments with decision trees and neural networks indicate that, in applications involving use of massive datasets, the simple approach of creating a committee of classifiers from disjoint partitions results in a fast and accurate classifier. The reduced complexity associated with the technique of producing random disjoint partitions makes it attractive for creating classifiers on extremely large datasets. We also propose a distributed version of pasting small votes, and achieve classification accuracy comparable to the sequential version as well as boosting. Distributed pasting of small votes is scalable and fast, even with very large datasets. We also highlight the inter-play between unstable or stable methods of learning classifiers and the diversity of classifiers.; A dataset is imbalanced if the classification categories are not approximately equally represented. Usually real-world datasets are predominantly composed of “normal” examples with only a small percentage of “abnormal” or “interesting” examples. Often the cost of misclassifying an abnormal (interesting) example as a normal example is much higher than the cost of the reverse error. Under-sampling of the majority (normal class) has been proposed as a good means of increasing the sensitivity of a classifier to the minority class. We show that a combination of over-sampling the minority (abnormal) class and under-sampling the majority (normal) class can achieve better classifier performance than only under-sampling the majority class. Our method of over-sampling the minority class involves creating synthetic minority class samples. Experiments are performed using C4.5, Ripper, and the Naive Bayes Classifier. We also show that our combination of over-sampling and under-sampling performs better than varying loss ratios in Ripper or by varying the class priors in the Naive Bayes Classifier: the methods that could directly handle skewed class distributions. We evaluate our approaches using ROC (AUC, ROC convex hull) analyses.