Python: Advanced Guide to Artificial Intelligence

Weighted log-likelihood

In the previous example, we have considered a single log-likelihood for both labeled and unlabeled samples:

This is equivalent to saying that we trust the unlabeled points just like the labeled ones. However, in some contexts, this assumption can lead to completely wrong estimations, as shown in the following graph:

Biased final Gaussian mixture configuration

In this case, the means and covariance matrices of both Gaussian distributions have been biased by the unlabeled points and the resulting density estimation is ...