Hands-On Unsupervised Learning with Python

Homogeneity score

The homogeneity score is complementary to the previous one and it's based on the assumption that a cluster must contain only samples having the same true label. It is defined as:

Analogously to the completeness score, when H(Y_true|Y_pred) → H(Y_true), it means that the assignments have no impact on the conditional entropy, hence the uncertainty is not reduced after the clustering (for example, every cluster contains samples belonging to all classes) and h → 0. Conversely, when H(Y_true|Y_pred) → 0, h → 1, because knowledge of the predictions has reduced the uncertainty about the true assignments and the clusters contain almost ...