Applications

Improving similarity learning classifiers

Learning a distance that fits the data properly will improve the accuracy of distance-based classifiers.

_images/ex_learning_nca.png

1-NN classification with euclidean distance (left), 1-NN classification after learning a distance (center) and the equivalent projection learned (right)

Dimensionality reduction

Many of the distance metric learning algorithms can learn projections onto low dimensional spaces. Dimensionality reduction improves the classifier eficiency, reduces overfitting and avoids problems such a the curse of dimensionality present in some similarity classifiers.

_images/ex_red_dim.png

The digits dataset (64 features) projected onto a plane with a distance metric learning algorithm.