Machine Learning
Machine Learning
Support Vector Machines for Pattern Classification (Advances in Pattern Recognition)
Support Vector Machines for Pattern Classification (Advances in Pattern Recognition)
Sparse least squares support vector training in the reduced empirical feature space
Pattern Analysis & Applications
Training of support vector machines with Mahalanobis kernels
ICANN'05 Proceedings of the 15th international conference on Artificial neural networks: formal models and their applications - Volume Part II
Feature selection based on kernel discriminant analysis
ICANN'06 Proceedings of the 16th international conference on Artificial Neural Networks - Volume Part II
Optimizing the kernel in the empirical feature space
IEEE Transactions on Neural Networks
Fast Sparse Approximation for Least Squares Support Vector Machine
IEEE Transactions on Neural Networks
Least-Squares Support Vector Machine Approach to Viral Replication Origin Prediction
INFORMS Journal on Computing
| Hi-index | 0.00 |
In our previous work, we have developed sparse least squares support vector machines (sparse LS SVMs) trained in the reduced empirical feature space, spanned by the independent training data selected by the Cholesky factorization. In this paper, we propose selecting the independent training data by forward selection based on linear discriminant analysis in the empirical feature space. Namely, starting from the empty set, we add a training datum that maximally separates two classes in the empirical feature space. To calculate the separability in the empirical feature space we use linear discriminant analysis (LDA), which is equivalent to kernel discriminant analysis in the feature space. If the matrix associated with the LDA is singular, we consider that the datum does not contribute to the class separation and permanently delete it from the candidates of addition. We stop the addition of data when the objective function of LDA does not increase more than the prescribed value. By computer experiments for two-class and multi-class problems we show that in most cases we can reduce the number of support vectors more than with the previous method.