Machine Learning
The impact of changing populations on classifier performance
KDD '99 Proceedings of the fifth ACM SIGKDD international conference on Knowledge discovery and data mining
Performance Analysis and Comparison of Linear Combiners for Classifier Fusion
Proceedings of the Joint IAPR International Workshop on Structural, Syntactic, and Statistical Pattern Recognition
A Theoretical and Experimental Analysis of Linear Combiners for Multiple Classifier Systems
IEEE Transactions on Pattern Analysis and Machine Intelligence
| Hi-index | 0.00 |
Most of what we know about multiple classifier systems is based on empirical findings, rather than theoretical results. Although there exist some theoretical results for simple and weighted averaging, it is difficult to gain an intuitive feel for classifier combination. In this paper we derive a bound on the region of the feature space in which the decision boundary can lie, for several methods of classifier combination using non-negative weights. This includes simple and weighted averaging of classifier outputs, and allows for a more intuitive understanding of the influence of the classifiers combined. We then apply this result to the design of a multiple logistic model for classifier combination in dynamic scenarios, and discuss its relevance to the concept of diversity amongst a set of classifiers.We consider the use of pairs of classifiers trained on label-swapped data, and deduce that although non-negative weights may be beneficial in stationary classification scenarios, for dynamic problems it is often necessary to use unconstrained weights for the combination.