Empirical Comparisons of Online Boosting Algorithms with Running Time

摘要

Boosting is an effective classifier combination method, which can improve classification performance of an unstable learning algorithm due to its theoretical performance guarantees and strong experimental results. However, the algorithm has been used mainly in batch mode, i.e., it requires the entire training set to be available at once and, in some cases, require random access to the data. Recently, Nikunj C.oza(2001) proved that some preliminary theoretical results and some empirical comparisons of the classification accuracies of online algorithms with their corresponding batch algorithms on many datasets. In this paper, we present online versions of some boosting methods that require only one pass through the training data. Specifically, we present theoretical and experimental evidence that our online algorithms succeed in this mirroring, often obtaining classification performance comparable to their batch counterparts in less time. We compare the online and batch algorithms experimentally in terms of running time.