Output details

Diversity in multiple classifier ensembles based on binary feature quantisation with application to face recognition

While the benefits from combining multiple classifiers in ensembles have been established by a large number of empirical studies, there are only a handful of papers examining the requirements for forming ensembles which outperform constituent classifiers. This paper, a major output from an EPSRC project (GR/S24831/01), offers a novel analysis of “diversity” which is recognized as one of the most important issues for the development of accurate multi-classifier systems. Our results strongly suggest that increased diversity among the constituent classifiers is not sufficient to guarantee performance improvements; rather a trade-off between participant classifiers’ accuracy and diversity is necessary to achieve maximum gains.