A combined approach to regularized linear combiner learning

Classifier combination is an important research area since they have a significant contribution to the accuracy. Even though simple fixed combination rules result in satisfactory performances, supervised combination learning surely has higher probability of better accuracy. Among supervised combination methods, linear combiner is one of the well-known methods. In this work, different types of linear combiners are examined in a unifying framework and a regularized least squares (LS) method is investigated for learning the combiner. Experiments are conducted on three different databases and results are examined. It is observed that, learning the weights is beneficial for higher accuracy then that obtained using simple sum rule.