You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
The $\ell^2$-norm is a commonly used metric in data science to compare similarity between distributions. However, if there is no spatial overlap between the features in the models being compared, then the $\ell^2$-norm cannot differentiate between their similarity, even though the distributions in the models could be very different. The Wasserstein metric (or earth mover's distance) quantifies the amount of work needed to transform one distribution into another, at the cost of higher computational expense.
Check out the animation below. The $\ell^2$-norm quickly saturates as the bulk of the two Gaussian blobs stop overlapping, while the Wasserstein metric increases at the same rate as their separation distance increases.
This is a simple toy example where the only change is that the two distributions increase in spatial separation. However, we could also imagine the width of the Gaussians increasing with distance (or changing into some other shape entirely). The $\ell^2$-norm would have no idea this was happening, while the Wasserstein metric would be able to quantify the difference this makes to the distributions' similarity.
About
Short Jupyter notebook demonstrating the advantage of using the Wasserstein distance over the L2 distance