关于Dunn指数为0的含义及K-means中指数变化的解读咨询

1. What does a Dunn Index value of 0 mean? #

Let’s start with the core formula: the Dunn Index = (smallest inter-cluster distance) / (largest intra-cluster distance).

A value of 0 tells you one unambiguous thing: the numerator (the smallest distance between any two different clusters) is 0. That translates directly to:

• At least two of your clusters have overlapping or identical samples. For example, there might be points that are exactly the same but got assigned to different clusters,
• Or one cluster is completely nested inside another—so the closest points between these two clusters are literally on top of each other,
• Or you’ve ended up with clusters that share a boundary where samples overlap, making their minimum inter-cluster distance zero.

In plain terms: a Dunn Index of 0 means your clustering result has at least two clusters that are not separated at all. They’re either overlapping or fully embedded in one another.

2. Interpreting conflicting Dunn Index and C-index results (k=5 vs k=6) #

First, let’s clarify the difference between these two metrics—they judge cluster quality through very different lenses:

• Dunn Index is a stickler for hard separation. It’s hyper-sensitive to even a single pair of overlapping points between any two clusters. It rewards clusters that are both tightly packed (small intra-cluster spread) and clearly separated (large inter-cluster gaps).
• C-index (I’m assuming you’re referring to the Calinski-Harabasz Index, a common "c-index" for clustering validation) focuses on variance ratios. It compares how spread out clusters are from each other versus how spread out points are within each cluster. It tends to favor clusterings where larger clusters are well-separated, and it’s way less sensitive to small, overlapping sub-clusters.

Now let’s unpack your specific scenario:

• When k=5, a Dunn Index of 0.05 means your clusters are weakly separated (the smallest gap between clusters is tiny compared to the biggest cluster’s spread), but there’s no complete overlap between any clusters.
• When k=6, the Dunn Index drops to 0: this happens because splitting one of the 5 clusters into two created at least two clusters with overlapping samples. Maybe you split a dense cluster into two sub-clusters that are so close together that some points are effectively in both, or the K-means algorithm assigned identical points to different clusters.
• Your C-index flagging k=5 as better makes perfect sense here. The C-index doesn’t penalize weak separation as harshly as the Dunn Index, and splitting into k=6 likely messed up the variance balance: the drop in intra-cluster variance from splitting wasn’t enough to offset the drop in inter-cluster variance. The small, overlapping clusters from k=6 don’t add meaningful separation, so the C-index correctly identifies k=5 as a more robust overall structure.

Key takeaway: #

The conflict boils down to what each metric considers "good clustering." The Dunn Index is a strict judge—even a tiny overlap ruins its score. The C-index takes a more holistic view of variance distribution. In your case, k=5 gives you a weakly but properly separated clustering, while k=6 introduces redundant, overlapping clusters that the Dunn Index rejects outright, even if the overall variance split is worse (per the C-index).

内容的提问来源于stack exchange，提问作者Gnturu Padmavathi