Adaptive selection of ensembles for imbalanced class distributions

Boolean combination (BC) techniques have been shown to efficiently integrate the responses of multiple diversified classifiers in the ROC space to improve the overall accuracy and reliability of pattern recognition systems. In practice, since class distributions are often imbalanced and change over time, the BC of classifiers, and thus selection of ensembles, should be adapted to reflect operational conditions. Although the impact on classification performance of imbalanced distributions may be addressed using ensemble-based techniques, this is difficult to observe from ROC curves. However, given a desired false positive rate and class imbalance, performing BC in the Precision-Recall Operating Characteristic (PROC) space with skewed data may lead to a higher level of performance. In this paper, an adaptive system is proposed that initially generates several PROC curves, each one from data with a different level of skew. Then, during operations, the class imbalance is periodically estimated, and used to approximate the most accurate BC of classifiers among operational points of these curves. Simulation results indicate that this approach maintains a high level of accuracy that is comparable to full Boolean re-combination (as required for a specific level of imbalance), but for a significantly lower computational cost.