Calculator Inputs and Chart Options
Provide paired labels and sizes, then configure the visual analysis.
Understanding Cluster Size Charts
Cluster size charts display how many observations belong to each discovered group. Bars make uneven assignments visible without requiring complex statistical review. They also help compare clustering runs across models, features, or datasets.
A balanced chart does not automatically prove that clustering is useful. Real populations may contain naturally small or dominant segments. The chart should support domain reasoning, validation metrics, and sample inspection.
Why Cluster Balance Matters
Cluster balance matters because extreme sizes can reveal weak model behavior. One cluster may absorb unrelated points after poor feature scaling. Tiny clusters may represent rare groups, noise, or unstable boundaries.
Compare cluster counts with silhouette scores and business relevance. Large differences deserve investigation before deployment or reporting. Strong conclusions require both numerical evidence and understandable cluster profiles.
Formula Used
The basic count formula is n sub k equals assigned observations. Percentage share divides each cluster count by the total count. The calculator multiplies that result by one hundred for display.
nk = number of observations assigned to cluster k
pk = nk / N
Percentagek = (nk / N) × 100
Mean = N / K
Imbalance Ratio = max(nk) / min(nk)
Entropy = -Σ pk ln(pk)
Mean size equals total observations divided by visible clusters. Standard deviation measures how far cluster sizes spread around that mean. The imbalance ratio divides the largest size by the smallest size.
Entropy summarizes how evenly observations are distributed among clusters. Normalized entropy approaches one when cluster shares are relatively even. Gini concentration rises when observations gather inside fewer clusters.
How to Use This Calculator
Enter cluster labels using commas, semicolons, or separate lines. Enter matching sizes using the same ordered structure. Every visible label must have one corresponding numeric size.
Choose sorting, thresholds, orientation, precision, and display options. Exclude noise when evaluating only modeled clusters. Keep noise included when reviewing operational assignment quality.
Press the calculation button after checking all entries. Results appear above the form for immediate comparison. The chart updates with summary metrics, shares, and selected annotations.
Chart Configuration Choices
Vertical bars suit short labels and standard reports. Horizontal bars improve readability when cluster names are lengthy. Percentage normalization supports comparisons across datasets having different totals.
A cumulative line creates a Pareto-style view of concentration. Log scaling helps when one cluster greatly exceeds smaller groups. However, log axes can hide absolute differences for casual readers.
Interpreting Advanced Metrics
Use the largest share to identify dominant cluster behavior. Review the smallest group before removing or merging it. Small groups sometimes contain valuable anomalies or specialized customer segments.
Inspect entropy and Gini together instead of using either metric alone. Entropy describes uncertainty across all observed shares during model evaluation. Gini supports quick concentration comparisons during practical model assessment.
Frequently Asked Questions
These answers explain common charting and interpretation decisions.
1. What does a cluster size bar chart show?
It shows observation counts assigned to each cluster. Taller bars represent larger groups. The chart quickly reveals dominance, sparsity, and possible assignment issues.
2. Should every cluster have a similar size?
No, equal sizes are not always expected. Natural populations often contain uneven segments. Evaluate balance alongside domain meaning, stability, and validation scores.
3. What is the imbalance ratio?
It divides the largest visible cluster by the smallest. Higher values indicate stronger size differences. Zero-sized clusters make this ratio undefined.
4. When should noise clusters remain included?
Include noise when reviewing density-based clustering or assignment coverage. Noise can reveal outliers and weak fit. Exclude it only for focused modeled-cluster comparisons.
5. Why use percentage normalization?
Percentages allow fair comparisons across datasets with different totals. They reveal relative cluster shares. Raw counts remain better for capacity planning.
6. What does normalized entropy indicate?
Values near one suggest relatively even cluster shares. Lower values indicate concentration within fewer clusters. Entropy does not measure semantic cluster quality.
7. When is a logarithmic scale useful?
Use it when cluster sizes differ by several magnitudes. Smaller bars become easier to inspect. Avoid it when audiences need direct absolute comparisons.
8. Can small clusters be removed automatically?
The minimum-size filter can hide smaller groups. Review them before filtering. Rare clusters may contain anomalies, risks, or valuable niches.
9. Can the results support model selection?
Yes, cluster sizes can support comparative evaluation. They should not decide alone. Combine them with stability, silhouette, separation, and practical usefulness.