CobLE: Confidence-Based Learning Ensembles

Combining information from a variety of sources greatly improves the classification accuracy compared with a single source. When the information sources are asynchronous (i.e., the combined feature set has missing values) and training data is limited, the accuracy of existing classification approaches are reduced. In this paper we present CobLE, an approach for creating an ensemble of classifiers. Each classifier operates on data from a single source and a "confidence" function is approximated for each classifier over its feature space. Classifier outputs are aggregated using weighted voting where the weight for each classifier is estimated from its confidence function. We present a theoretical analysis and extensive experimental results demonstrating significant improvement over existing ensemble learning and data fusion approaches, especially with asynchronous data sources. We also present a thorough evaluation of the effects of CobLE's internal parameters on performance.