Fig. 1

Experiments on feature weight estimation on Fermat’s Spiral. a Each class of 200 samples is labeled by a different color. To test the accuracy of feature weighting by LHDA, artificial noisy features of various dimensions (0 to 1000) were added to the dataset. The first two features completely determine the labels of the synthetic samples, while other features are redundant noises. These results are consistent with the data setting scheme. Estimated feature weights are plotted for noisy features of dimensions (b) 100; (c) 600; and (d) 1000