Cluster Validity

SSE

Using the elbow method to find the number of clusters

Statistical Framework for Cluster Validity

• Randomy generate data, and apply your clustering solution

External Methods

You need some ground-truth labels

• Purity of cluster $i$: $\displaystyle \max_{j}\left[\frac{m_{ij}}{\sum_k m_{ik}}\right]$

• Entropy of cluster $i$: $-\displaystyle \sum_{j=1}^C\frac{m_{ij}}{\sum_k m_{ik}}\log_2 \frac{m_{ij}}{\sum_k m_{ik}}$

  Lower entropy indicates higher purity.