3. Dimensionality reduction

The diversity of algorithms used in dimensionality reduction is due first and foremost to the fact that the objectives of dimensionality reduction vary widely.

If we use this technique simply for visualization purposes, to understand the distribution of a data set, we'll naturally look for an algorithm that best preserves the local or global structure of the data, taking as a reference, for example, the respective distances between each example. If, on the other hand, such a dimensionality reduction is used to feed a classification algorithm, the algorithm will be asked to retain the most discriminating variables for the classification that follows.