Transductive methods for distributed ensemble classification

We consider ensemble classification for the case when there is no common labeled training data for designing the function which aggregates individual classifier decisions. We dub this problem distributed ensemble classification, addressing e.g. when individual classifiers are trained on different (e.g. proprietary, legacy) databases or operate (perhaps remotely) on different sensing modalities. Typically, fixed, principled (untrained) rules of classifier combination such as voting methods are used in this case for aggregating decisions. Alternatively, we take a transductive approach, optimizing the combining rule for an objective function measured on the unlabeled batch of test data. We propose specific maximum likelihood (ML) objectives that are shown to yield well-known forms of aggregation, albeit with iterative, EM-based adjustment to account for possible mismatch between the class priors used by individual classifiers and those reflected in the new data batch. We also propose an information-theoretic method which outperforms the ML methods and addresses some problem instances where the ML methods are not applicable. On benchmark data from the UC Irvine machine learning repository, all our methods give improvements in accuracy over the use of fixed rules when there is prior mismatch.