Search code examples
pythonscikit-learncluster-analysis

Difference between NMI and AMI scores

Can someone explain the difference between two normalized versions of measurements (NMI and AMI) that measure the agreement of the two assignments, ignoring permutations.

Let consider this code: 

from sklearn import metrics
labels_true = [0, 0, 0, 1, 1, 1]
labels_pred = [0, 0, 1, 1, 2, 2]

# AMI score: 
score_ami = metrics.adjusted_mutual_info_score(labels_true, labels_pred)  
print(score_ami)

# NMI Score
score_nmi = metrics.normalized_mutual_info_score(labels_true, labels_pred)  
print(score_nmi)
Solution

  • Adjusted Mutual Information is rescaled such that a random clustering has score 0.

    With NMI, even randomly shuffled labels will get a positive score.