We compare experimentally the performance of three approaches to ensemble-based classification on general multi-class datasets. These are the methods of random forest, error-correcting output codes (ECOC) and ECOC enhanced by the use of bootstrapping and class-separability weighting (ECOC-BW). These experiments suggest that ECOC-BW yields better generalisation performance than either random forest or unmodified ECOC. A bias-variance analysis indicates that ECOC benefits from reduced bias, when compared to random forest, and that ECOC-BW benefits additionally from reduced variance. One disadvantage of ECOC-based algorithms, however, when compared with random forest, is that they impose a greater computational demand leading to longer training times. © 2011 Springer-Verlag.
CITATION STYLE
Smith, R. S., Bober, M., & Windeatt, T. (2011). A comparison of random forest with ECOC-based classifiers. In Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) (Vol. 6713 LNCS, pp. 207–216). https://doi.org/10.1007/978-3-642-21557-5_23
Mendeley helps you to discover research relevant for your work.