Evaluate cluster compactness and separation using flexible metrics, scaling methods, detailed diagnostics, visual charts, model comparisons, and exportable machine learning reports instantly online today.
| Cluster | Size | Calculated Representative | Used Representative | Scatter Si | Worst Match | Maximum Ri | DB Contribution |
|---|
The index compares cluster scatter against cluster separation. Each cluster uses its worst pairwise similarity. The final score averages those worst similarities.
| Cluster | Feature 1 | Feature 2 |
|---|---|---|
| A | 1.0 | 1.1 |
| A | 1.3 | 0.9 |
| B | 5.2 | 5.0 |
| B | 4.8 | 5.4 |
| C | 8.0 | 1.8 |
| C | 8.4 | 1.4 |
Lower scores generally represent compact and separated clusters. The measure supports fast comparisons across matching datasets. It remains easy to calculate and explain.
The result depends on scaling and distance choices. Outliers can increase scatter and distort comparisons. Irregular cluster shapes may receive misleading scores.
No universal score guarantees a correct clustering solution. Compare models under identical preprocessing and feature selections. Review silhouette and Calinski-Harabasz scores alongside it.
A lower score usually indicates compact, separated clusters. It reflects smaller scatter relative to centroid separation. Always compare scores under identical data preparation.
Zero is the theoretical lower bound. It requires clusters with no internal scatter. Real datasets usually produce positive values.
No universal threshold fits every dataset. Feature scales and geometry strongly affect results. Relative model comparison is usually more reliable.
Standardisation helps when feature ranges differ greatly. Otherwise large-scale features dominate calculated distances. Use domain knowledge before transforming meaningful scales.
Yes, medoids are available for robust representatives. They are actual observations inside each cluster. Standard Davies-Bouldin calculations normally use arithmetic centroids.
Different clusters can share identical representatives. Their separation then equals zero. The pairwise similarity becomes infinite or undefined.
Yes, save each completed calculation as a model. The comparison table ranks lower finite scores first. Keep preprocessing and features consistent during comparison.
Yes, the calculator supports cosine distance. Zero vectors require careful handling. Euclidean distance remains the conventional default choice.
No single internal metric captures every cluster property. Inspect visualisations and alternative validation measures too. Domain usefulness remains the final evaluation standard.
Important Note: All the Calculators listed in this site are for educational purpose only and we do not guarentee the accuracy of results. Please do consult with other sources as well.