Le fléau d’un trop grand nombre de variables

Adding features feels helpful, but past a point each new column makes your data sparser and your model harder to train. 😬

This squeeze is the curse of dimensionality: as dimensions grow, the space balloons and your points scatter far apart.

In very high dimensions almost every pair of points sits roughly the same distance apart, so distance-based methods stop telling things apart.

To keep the same density, the rows you need grow exponentially with features. Real datasets never have that many rows, so space stays mostly empty.

With many columns and few rows, a model can memorize noise instead of signal. That overfitting looks great in training and fails on new data.

Many features quietly repeat each other, like height in cm and height in inches. This redundancy adds cost without adding new information.

Every extra column carries a little measurement noise. Stack enough of them and the noise can drown out the few features that truly matter.

More dimensions mean more memory and longer training. Wide tables make even simple models slow and awkward to tune.

You can plot two or three dimensions, but not fifty. High-dimensional data is nearly impossible to visualize or reason about directly.

You can drop weak columns with feature selection, or combine columns into fewer new ones with feature extraction like PCA.

PCA compresses many correlated features into a handful of new axes, keeping most of the information while cutting the dimension count.

Think about what really breaks as dimensions grow.

Too many features bring the curse of dimensionality: sparse data, fuzzy distances, and overfitting. Reducing dimensions, often with PCA, fixes it. 🎯