I am clustering time-series datasets which are not labeled (No Ground truth) and I want to measure the quality of the clusters. Could you please suggest any Clustering performance evaluation methods that can be used in time-series clustering ??
1 Answer
Clustering evaluation metrics basic goal is to measure the similarity within each cluster and dissimilarity among clusters; if a clustering algorithm has achieved those two things in an acceptable degree than it has performed well.
The most commonly used evaluation metrics are:
-Silhouette Coefficient - it is the most popular one for time-series clustering (implementation for python can be found in tslearn.clustering package)
-Davies-Bouldin Index (implemenation for python can be found in sklearn package)
-Dunn's index
You might find the following articles useful sources to help you understand those metrics:
- "Evaluation Metrics for Clustering" in Medium
- Article with presentation of Evaluation Metric for Supervised and Unsupervised learning
- Explanation of Dunn index and DB index
- An example using Silhouette score on time-series clustering
Finally, a more complete collection of clustering evaluation measures can be found in this review.