Supervised (a.k.a. Task-specific) dimensionality reduction

However, while unsupervised dimensionality reduction is indispensable as an “in-general” tool, sometimes, we require a more intelligent approach for the task at hand. Sometimes, instead of focusing on preserving the as much variance as possible in a lower-dimensional projection, we are better served by picking and choosing features/dimensions. A dummy example of such an instance is shown in the figure above, where there are 2 classes in a high-dimensional dataset (10,000 dimensions). While we CAN certainly work in such a high-dimensional space, it would be more prudent to try and distill the essence of these data to a lower-dimensional space for computational efficiency (not to mention the fact that many learning algorithms actually perform rather poorly on high-dimensional data). <a href="https://medium.com/@abuzar_mahmood/supervised-a-k-a-task-specific-dimensionality-reduction-c9325bdb08a2">Read More</a>