S. Singh, V. Kumar, and M. Singh (UK)
Subspace Classifiers, Classification Complexity, Feature Space Partitioning.
Multiresolution estimates of classification complexity estimate the relative ease with which multivariate data belonging to multiple classes can be separated by non linear boundaries in high dimensional spaces. In this paper we propose the concept of using multiple classifiers in feature subspaces that are generated by feature space partitioning. We find that the advantage gained by training multiple classifiers for a given data set is far greater than the disadvantage of having less number of samples in each feature subspace to train them. In this paper we take a number of data sets from the UCI repository and show the classification advantage gained by using multiple subspace classifiers in parallel. We also demonstrate that the multi-resolution estimates of classification complexity correlate well with this classification performance averaged across all subspaces.
Important Links:
Go Back
