Instance-Based Learning Algorithms
Machine Learning
Reduction Techniques for Instance-BasedLearning Algorithms
Machine Learning
On Issues of Instance Selection
Data Mining and Knowledge Discovery
A Unified Bias-Variance Decomposition for Zero-One and Squared Loss
Proceedings of the Seventeenth National Conference on Artificial Intelligence and Twelfth Conference on Innovative Applications of Artificial Intelligence
Machine learning with data dependent hypothesis classes
The Journal of Machine Learning Research
Approximation Algorithms for the Class Cover Problem
Annals of Mathematics and Artificial Intelligence
Random Graphs for Statistical Pattern Recognition
Random Graphs for Statistical Pattern Recognition
Statistical Comparisons of Classifiers over Multiple Data Sets
The Journal of Machine Learning Research
Presupervised and post-supervised prototype classifier design
IEEE Transactions on Neural Networks
Hi-index | 0.00 |
This paper describes an instance based classifier, the randomised sphere covering classifier (αRSC), that reduces the training data set size without loss of accuracy when compared to nearest neighbour classifiers. The motivation for developing this algorithm is the desire to have a non-deterministic, fast, instance based classifier that performs well in isolation but is also ideal for use with ensembles. Essentially we trade off decreased testing time for increased training time whilst retaining the simple and intuitive nature of instance based classifiers. We use twenty four benchmark datasets from UCI repository for evaluation. The first set of experiments demonstrate the basic benefits of sphere covering. We show that there is no significant difference in accuracy between the basic αRSC algorithm and a nearest neighbour classifier, even though αRSC compresses the data by up to 75%. We then describe a pruning algorithm that removes spheres that contain α or fewer training instances. The second set of experiments demonstrate that when we set the α parameter through cross validation, the resulting αRSC algorithm outperforms several well known classifiers when compared using the Friedman rank sum test. Thirdly, we highlight the benefits of pruning with a bias/variance decomposition. Finally, we discuss why the randomisation inherent in αRSC makes them an ideal ensemble component and outline our future direction.