Chapter map

1. 1. Definition
2. 2. Two clusters
3. 3. BH two
4. 4. Three clusters
5. 5. BH three
6. 6. Four clusters
7. 7. BH four
8. 8. First drop
9. 9. Second drop
10. 10. Selection
11. 11. Monotonicity
12. 12. Cluster weighting
13. 13. Singleton
14. 14. Restarts
15. 15. Range
16. 16. Claim

Visit the How Mathematics Works hub for the wider series and its connections to education, reasoning, science, computing, design and careers.

---

1. Definition

This is where modelling choices begin to affect evidence. Use BH(k)=k^-1 sum W_i/n_i.

A concrete situation

Each cluster contributes one mean squared dispersion.

What the mathematics does not license

Other summaries may write W/k differently. That distinction matters because a correct formula can still answer the wrong operational question.

Try it, check it, transfer it

State formula. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

2. Two clusters

The central mathematical idea is concrete. Use mean dispersions 5 and 3.

A concrete situation

The sum is 8.

What the mathematics does not license

Both use the same distance and scaling. The habit of checking assumptions is one of the practical benefits of learning mathematics.

Try it, check it, transfer it

Record values. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

3. BH two

Here the geometry becomes operational. Divide 8 by 2.

A concrete situation

BH2=4.

What the mathematics does not license

Squared-distance units remain. The same reasoning helps students move from an exercise to a real dataset.

Try it, check it, transfer it

Show step. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

4. Three clusters

This step deserves a careful pause. Use 2, 2.5 and 1.5.

A concrete situation

The sum is 6.

What the mathematics does not license

Small clusters receive equal cluster weight. That distinction matters because a correct formula can still answer the wrong operational question.

Try it, check it, transfer it

Record values. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

5. BH three

The useful question is not only how to calculate, but what the calculation assumes. Divide 6 by 3.

A concrete situation

BH3=2.

What the mathematics does not license

The index falls by 2. The habit of checking assumptions is one of the practical benefits of learning mathematics.

Try it, check it, transfer it

Show step. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

6. Four clusters

A small example makes the mechanism visible. Use 1.8,1.7,1.4,1.3.

A concrete situation

The sum is 6.2.

What the mathematics does not license

Every candidate needs comparable fits. The same reasoning helps students move from an exercise to a real dataset.

Try it, check it, transfer it

Record values. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

7. BH four

This is where modelling choices begin to affect evidence. Divide 6.2 by 4.

A concrete situation

BH4=1.55.

What the mathematics does not license

The index falls again. That distinction matters because a correct formula can still answer the wrong operational question.

Try it, check it, transfer it

Show step. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

8. First drop

The central mathematical idea is concrete. Compute 4-2.

A concrete situation

Drop into k=3 is 2.

What the mathematics does not license

Declare destination indexing. The habit of checking assumptions is one of the practical benefits of learning mathematics.

Try it, check it, transfer it

Record drop. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

9. Second drop

Here the geometry becomes operational. Compute 2-1.55.

A concrete situation

Drop into k=4 is .45.

What the mathematics does not license

Round after subtraction. The same reasoning helps students move from an exercise to a real dataset.

Try it, check it, transfer it

Record drop. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

10. Selection

This step deserves a careful pause. Compare 2 with .45.

A concrete situation

Largest drop points to k=3.

What the mathematics does not license

Software conventions can label the transition differently. That distinction matters because a correct formula can still answer the wrong operational question.

Try it, check it, transfer it

State rule. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

11. Monotonicity

The useful question is not only how to calculate, but what the calculation assumes. Expect dispersion to decline as k rises.

A concrete situation

Raw minimum alone would overfit.

What the mathematics does not license

Use changes, not the terminal value. The habit of checking assumptions is one of the practical benefits of learning mathematics.

Try it, check it, transfer it

Plot sequence. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

12. Cluster weighting

A small example makes the mechanism visible. Compare equal cluster and point weighting.

A concrete situation

Small clusters can influence BH strongly.

What the mathematics does not license

State chosen formula. The same reasoning helps students move from an exercise to a real dataset.

Try it, check it, transfer it

Run both. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

13. Singleton

This is where modelling choices begin to affect evidence. Let n_i=1 and W_i=0.

A concrete situation

A singleton contributes zero dispersion.

What the mathematics does not license

Outliers can be isolated cheaply. That distinction matters because a correct formula can still answer the wrong operational question.

Try it, check it, transfer it

Flag singleton. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

14. Restarts

The central mathematical idea is concrete. Repeat algorithm initialisations.

A concrete situation

Local minima change W_i.

What the mathematics does not license

One run can invent an elbow. The habit of checking assumptions is one of the practical benefits of learning mathematics.

Try it, check it, transfer it

Report spread. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

15. Range

Here the geometry becomes operational. Extend candidate k values.

A concrete situation

The largest visible drop can move.

What the mathematics does not license

Range is part of evidence. The same reasoning helps students move from an exercise to a real dataset.

Try it, check it, transfer it

Declare range. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

16. Claim

This step deserves a careful pause. Name W_i,n_i,formula,sequence and rule.

A concrete situation

Teaching transition favours three.

What the mathematics does not license

Do not guarantee usefulness. That distinction matters because a correct formula can still answer the wrong operational question.

Try it, check it, transfer it

Bound conclusion. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

A seven-day practice lab

1. Day 1: define the decision, observation unit, X variables and Y or reference condition.
2. Day 2: reproduce the smallest worked example and check every dimension, unit and sign.
3. Day 3: build a transparent baseline and write what it assumes.
4. Day 4: fit the main method inside a training-only pipeline; record every parameter.
5. Day 5: challenge one assumption with a deliberately difficult dataset or operating regime.
6. Day 6: compare validation metrics with the practical workload created by errors.
7. Day 7: present one supported conclusion, one unresolved uncertainty and one safe next experiment.

Questions for a student, parent or mentor

• Which part of the model is a mathematical definition, and which part is a choice?
• What small calculation would catch a sign, scale or dimension error?
• What evidence comes from training, validation and the untouched test?
• What benign condition could look like a fault or prediction failure?
• How would the answer change for another response, asset, school project or time period?
• What should a human review before acting on the score?

A method-specific audit before deployment

The Ball–Hall index averages within-cluster dispersion across clusters. In one common form, BH(k)=(1/k) times the sum over clusters of W_i/n_i, where W_i is the squared distance total around cluster i and n_i is its membership count. Smaller values mean tighter clusters, but adding clusters tends to reduce the index, so the useful signal is a pronounced change between successive levels.

For k=2, suppose the two cluster mean squared dispersions are 5 and 3. Then BH(2)=(5+3)/2=4. For k=3, suppose the values are 2, 2.5 and 1.5, giving BH(3)=2.

For k=4, suppose the four values are 1.8, 1.7, 1.4 and 1.3, giving BH(4)=1.55. The drop from two to three clusters is 4-2=2, while the next drop is 2-1.55=0.45. Under the declared destination convention, the largest drop points to three clusters.

Software can implement the successive-level rule with different indexing or use a slope or second-difference convention. Cluster weighting matters too: averaging W_i/n_i gives each cluster equal influence regardless of size. Report the full sequence and the exact selection rule instead of publishing only a winning k.

---

Frequently asked questions

What is the Ball–Hall index?

An average of within-cluster dispersion contributions across the clusters in a partition.

Why not simply choose its smallest value?

Dispersion usually declines as clusters are added, so a successive-change rule looks for a pronounced improvement.

Which cluster count wins?

Under the stated destination convention, the count reached by the largest successive drop.

Why do small clusters matter?

Equal cluster averaging can give a small cluster as much influence as a large cluster.

What happens with a singleton?

Its within-cluster dispersion is zero, which can make outlier isolation look attractive.

What must be reported?

Formula, W values, n values, weighting, index sequence, difference convention, candidate range, algorithm, scaling and random starts.

---

Primary source and fact-check note

Geoffrey H. Ball and David J. Hall, ISODATA, A Novel Method of Data Analysis and Pattern Classification, Stanford Research Institute Technical Report, 1965; selection convention cross-checked in the current NbClust reference manual. Definitions and core claims in this article were checked against this primary publication. Calculations shown here were independently recomputed. Examples are educational illustrations; they are not reported industrial performance results.

Useful next reading

• Krzanowski–Lai dimension-corrected changes
• Calinski–Harabasz variance ratio
• Gap statistic and reference distributions
• How AI finds patterns without labels
• Bukit Timah Mathematics learning routes

Keep the options open. Mathematics education supports clearer questions, transferable problem-solving skills and informed choices across science, engineering, computing, AI, statistics, design and many other careers. It does not guarantee intelligence, income, admission or employment; those outcomes depend on far more than one subject or model.

Extended laboratory: cluster-validity evidence and responsible use

Data audit

Freeze observations, features, units, missing-data rules, transformations, standardisation, sample size and random seeds.

Partition audit

Use comparable clustering algorithms, distances, candidate ranges, stopping rules and random-start budgets.

Formula audit

Show every scatter norm, density ratio, dispersion contribution, pair count, distance mean, likelihood, parameter penalty and entropy term.

Convention audit

Name ordered or unordered pairs, density radius, destination indexing, weighting, logarithm base and higher-or-lower score direction.

Uncertainty audit

Repeat starts, samples, scales, distances, candidate ranges, density radii and covariance parameterisations.

Failure audit

Test singletons, duplicated points, zero denominators, crowded midpoints, imbalanced groups, local optima and singular mixtures.

Meaning audit

Compare internal numerical evidence with stability, prediction and domain evidence without treating clusters as automatic categories.

Final claim

State the exact geometric or probabilistic evidence and stop before causal, semantic, career or guaranteed-performance claims.

Mechanism workshop: calculate, challenge and transfer

Workshop 1: Definition

Begin with this cluster-validity mechanism: Use BH(k)=k^-1 sum W_i/n_i. Recreate the concrete case: Each cluster contributes one mean squared dispersion. Write candidate k, W_i, n_i, cluster contributions, BH values, successive drops, indexing rule and selected transition with identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Other summaries may write W/k differently. Complete the transfer step: State formula. Change exactly one cluster size, pair count, distance, feature scale, scatter norm, density radius, midpoint count, dispersion value, candidate range, likelihood, responsibility, entropy value, covariance model, initialisation or seed. Explain which mathematical object changes, which score moves and which interpretation must weaken. Keep compactness, separation, overlap, stability, uncertainty and practical usefulness on separate notebook lines. Explain once to a student and once to a parent. Add a counterexample involving a singleton, duplicated point, zero denominator, density bridge, unequal group size, local optimum or overlapping Gaussian components. State the extra evidence needed to distinguish explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 2: Two clusters

Begin with this cluster-validity mechanism: Use mean dispersions 5 and 3. Recreate the concrete case: The sum is 8. Write candidate k, W_i, n_i, cluster contributions, BH values, successive drops, indexing rule and selected transition with identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Both use the same distance and scaling. Complete the transfer step: Record values. Change exactly one cluster size, pair count, distance, feature scale, scatter norm, density radius, midpoint count, dispersion value, candidate range, likelihood, responsibility, entropy value, covariance model, initialisation or seed. Explain which mathematical object changes, which score moves and which interpretation must weaken. Keep compactness, separation, overlap, stability, uncertainty and practical usefulness on separate notebook lines. Explain once to a student and once to a parent. Add a counterexample involving a singleton, duplicated point, zero denominator, density bridge, unequal group size, local optimum or overlapping Gaussian components. State the extra evidence needed to distinguish explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 3: BH two

Begin with this cluster-validity mechanism: Divide 8 by 2. Recreate the concrete case: BH2=4. Write candidate k, W_i, n_i, cluster contributions, BH values, successive drops, indexing rule and selected transition with identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Squared-distance units remain. Complete the transfer step: Show step. Change exactly one cluster size, pair count, distance, feature scale, scatter norm, density radius, midpoint count, dispersion value, candidate range, likelihood, responsibility, entropy value, covariance model, initialisation or seed. Explain which mathematical object changes, which score moves and which interpretation must weaken. Keep compactness, separation, overlap, stability, uncertainty and practical usefulness on separate notebook lines. Explain once to a student and once to a parent. Add a counterexample involving a singleton, duplicated point, zero denominator, density bridge, unequal group size, local optimum or overlapping Gaussian components. State the extra evidence needed to distinguish explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 4: Three clusters

Begin with this cluster-validity mechanism: Use 2, 2.5 and 1.5. Recreate the concrete case: The sum is 6. Write candidate k, W_i, n_i, cluster contributions, BH values, successive drops, indexing rule and selected transition with identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Small clusters receive equal cluster weight. Complete the transfer step: Record values. Change exactly one cluster size, pair count, distance, feature scale, scatter norm, density radius, midpoint count, dispersion value, candidate range, likelihood, responsibility, entropy value, covariance model, initialisation or seed. Explain which mathematical object changes, which score moves and which interpretation must weaken. Keep compactness, separation, overlap, stability, uncertainty and practical usefulness on separate notebook lines. Explain once to a student and once to a parent. Add a counterexample involving a singleton, duplicated point, zero denominator, density bridge, unequal group size, local optimum or overlapping Gaussian components. State the extra evidence needed to distinguish explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 5: BH three

Begin with this cluster-validity mechanism: Divide 6 by 3. Recreate the concrete case: BH3=2. Write candidate k, W_i, n_i, cluster contributions, BH values, successive drops, indexing rule and selected transition with identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: The index falls by 2. Complete the transfer step: Show step. Change exactly one cluster size, pair count, distance, feature scale, scatter norm, density radius, midpoint count, dispersion value, candidate range, likelihood, responsibility, entropy value, covariance model, initialisation or seed. Explain which mathematical object changes, which score moves and which interpretation must weaken. Keep compactness, separation, overlap, stability, uncertainty and practical usefulness on separate notebook lines. Explain once to a student and once to a parent. Add a counterexample involving a singleton, duplicated point, zero denominator, density bridge, unequal group size, local optimum or overlapping Gaussian components. State the extra evidence needed to distinguish explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 6: Four clusters

Begin with this cluster-validity mechanism: Use 1.8,1.7,1.4,1.3. Recreate the concrete case: The sum is 6.2. Write candidate k, W_i, n_i, cluster contributions, BH values, successive drops, indexing rule and selected transition with identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Every candidate needs comparable fits. Complete the transfer step: Record values. Change exactly one cluster size, pair count, distance, feature scale, scatter norm, density radius, midpoint count, dispersion value, candidate range, likelihood, responsibility, entropy value, covariance model, initialisation or seed. Explain which mathematical object changes, which score moves and which interpretation must weaken. Keep compactness, separation, overlap, stability, uncertainty and practical usefulness on separate notebook lines. Explain once to a student and once to a parent. Add a counterexample involving a singleton, duplicated point, zero denominator, density bridge, unequal group size, local optimum or overlapping Gaussian components. State the extra evidence needed to distinguish explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 7: BH four

Begin with this cluster-validity mechanism: Divide 6.2 by 4. Recreate the concrete case: BH4=1.55. Write candidate k, W_i, n_i, cluster contributions, BH values, successive drops, indexing rule and selected transition with identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: The index falls again. Complete the transfer step: Show step. Change exactly one cluster size, pair count, distance, feature scale, scatter norm, density radius, midpoint count, dispersion value, candidate range, likelihood, responsibility, entropy value, covariance model, initialisation or seed. Explain which mathematical object changes, which score moves and which interpretation must weaken. Keep compactness, separation, overlap, stability, uncertainty and practical usefulness on separate notebook lines. Explain once to a student and once to a parent. Add a counterexample involving a singleton, duplicated point, zero denominator, density bridge, unequal group size, local optimum or overlapping Gaussian components. State the extra evidence needed to distinguish explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 8: First drop

Begin with this cluster-validity mechanism: Compute 4-2. Recreate the concrete case: Drop into k=3 is 2. Write candidate k, W_i, n_i, cluster contributions, BH values, successive drops, indexing rule and selected transition with identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Declare destination indexing. Complete the transfer step: Record drop. Change exactly one cluster size, pair count, distance, feature scale, scatter norm, density radius, midpoint count, dispersion value, candidate range, likelihood, responsibility, entropy value, covariance model, initialisation or seed. Explain which mathematical object changes, which score moves and which interpretation must weaken. Keep compactness, separation, overlap, stability, uncertainty and practical usefulness on separate notebook lines. Explain once to a student and once to a parent. Add a counterexample involving a singleton, duplicated point, zero denominator, density bridge, unequal group size, local optimum or overlapping Gaussian components. State the extra evidence needed to distinguish explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 9: Second drop

Begin with this cluster-validity mechanism: Compute 2-1.55. Recreate the concrete case: Drop into k=4 is .45. Write candidate k, W_i, n_i, cluster contributions, BH values, successive drops, indexing rule and selected transition with identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Round after subtraction. Complete the transfer step: Record drop. Change exactly one cluster size, pair count, distance, feature scale, scatter norm, density radius, midpoint count, dispersion value, candidate range, likelihood, responsibility, entropy value, covariance model, initialisation or seed. Explain which mathematical object changes, which score moves and which interpretation must weaken. Keep compactness, separation, overlap, stability, uncertainty and practical usefulness on separate notebook lines. Explain once to a student and once to a parent. Add a counterexample involving a singleton, duplicated point, zero denominator, density bridge, unequal group size, local optimum or overlapping Gaussian components. State the extra evidence needed to distinguish explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 10: Selection

Begin with this cluster-validity mechanism: Compare 2 with .45. Recreate the concrete case: Largest drop points to k=3. Write candidate k, W_i, n_i, cluster contributions, BH values, successive drops, indexing rule and selected transition with identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Software conventions can label the transition differently. Complete the transfer step: State rule. Change exactly one cluster size, pair count, distance, feature scale, scatter norm, density radius, midpoint count, dispersion value, candidate range, likelihood, responsibility, entropy value, covariance model, initialisation or seed. Explain which mathematical object changes, which score moves and which interpretation must weaken. Keep compactness, separation, overlap, stability, uncertainty and practical usefulness on separate notebook lines. Explain once to a student and once to a parent. Add a counterexample involving a singleton, duplicated point, zero denominator, density bridge, unequal group size, local optimum or overlapping Gaussian components. State the extra evidence needed to distinguish explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 11: Monotonicity

Begin with this cluster-validity mechanism: Expect dispersion to decline as k rises. Recreate the concrete case: Raw minimum alone would overfit. Write candidate k, W_i, n_i, cluster contributions, BH values, successive drops, indexing rule and selected transition with identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Use changes, not the terminal value. Complete the transfer step: Plot sequence. Change exactly one cluster size, pair count, distance, feature scale, scatter norm, density radius, midpoint count, dispersion value, candidate range, likelihood, responsibility, entropy value, covariance model, initialisation or seed. Explain which mathematical object changes, which score moves and which interpretation must weaken. Keep compactness, separation, overlap, stability, uncertainty and practical usefulness on separate notebook lines. Explain once to a student and once to a parent. Add a counterexample involving a singleton, duplicated point, zero denominator, density bridge, unequal group size, local optimum or overlapping Gaussian components. State the extra evidence needed to distinguish explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 12: Cluster weighting

Begin with this cluster-validity mechanism: Compare equal cluster and point weighting. Recreate the concrete case: Small clusters can influence BH strongly. Write candidate k, W_i, n_i, cluster contributions, BH values, successive drops, indexing rule and selected transition with identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: State chosen formula. Complete the transfer step: Run both. Change exactly one cluster size, pair count, distance, feature scale, scatter norm, density radius, midpoint count, dispersion value, candidate range, likelihood, responsibility, entropy value, covariance model, initialisation or seed. Explain which mathematical object changes, which score moves and which interpretation must weaken. Keep compactness, separation, overlap, stability, uncertainty and practical usefulness on separate notebook lines. Explain once to a student and once to a parent. Add a counterexample involving a singleton, duplicated point, zero denominator, density bridge, unequal group size, local optimum or overlapping Gaussian components. State the extra evidence needed to distinguish explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 13: Singleton

Begin with this cluster-validity mechanism: Let n_i=1 and W_i=0. Recreate the concrete case: A singleton contributes zero dispersion. Write candidate k, W_i, n_i, cluster contributions, BH values, successive drops, indexing rule and selected transition with identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Outliers can be isolated cheaply. Complete the transfer step: Flag singleton. Change exactly one cluster size, pair count, distance, feature scale, scatter norm, density radius, midpoint count, dispersion value, candidate range, likelihood, responsibility, entropy value, covariance model, initialisation or seed. Explain which mathematical object changes, which score moves and which interpretation must weaken. Keep compactness, separation, overlap, stability, uncertainty and practical usefulness on separate notebook lines. Explain once to a student and once to a parent. Add a counterexample involving a singleton, duplicated point, zero denominator, density bridge, unequal group size, local optimum or overlapping Gaussian components. State the extra evidence needed to distinguish explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 14: Restarts

Begin with this cluster-validity mechanism: Repeat algorithm initialisations. Recreate the concrete case: Local minima change W_i. Write candidate k, W_i, n_i, cluster contributions, BH values, successive drops, indexing rule and selected transition with identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: One run can invent an elbow. Complete the transfer step: Report spread. Change exactly one cluster size, pair count, distance, feature scale, scatter norm, density radius, midpoint count, dispersion value, candidate range, likelihood, responsibility, entropy value, covariance model, initialisation or seed. Explain which mathematical object changes, which score moves and which interpretation must weaken. Keep compactness, separation, overlap, stability, uncertainty and practical usefulness on separate notebook lines. Explain once to a student and once to a parent. Add a counterexample involving a singleton, duplicated point, zero denominator, density bridge, unequal group size, local optimum or overlapping Gaussian components. State the extra evidence needed to distinguish explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 15: Range

Begin with this cluster-validity mechanism: Extend candidate k values. Recreate the concrete case: The largest visible drop can move. Write candidate k, W_i, n_i, cluster contributions, BH values, successive drops, indexing rule and selected transition with identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Range is part of evidence. Complete the transfer step: Declare range. Change exactly one cluster size, pair count, distance, feature scale, scatter norm, density radius, midpoint count, dispersion value, candidate range, likelihood, responsibility, entropy value, covariance model, initialisation or seed. Explain which mathematical object changes, which score moves and which interpretation must weaken. Keep compactness, separation, overlap, stability, uncertainty and practical usefulness on separate notebook lines. Explain once to a student and once to a parent. Add a counterexample involving a singleton, duplicated point, zero denominator, density bridge, unequal group size, local optimum or overlapping Gaussian components. State the extra evidence needed to distinguish explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 16: Claim

Begin with this cluster-validity mechanism: Name W_i,n_i,formula,sequence and rule. Recreate the concrete case: Teaching transition favours three. Write candidate k, W_i, n_i, cluster contributions, BH values, successive drops, indexing rule and selected transition with identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Do not guarantee usefulness. Complete the transfer step: Bound conclusion. Change exactly one cluster size, pair count, distance, feature scale, scatter norm, density radius, midpoint count, dispersion value, candidate range, likelihood, responsibility, entropy value, covariance model, initialisation or seed. Explain which mathematical object changes, which score moves and which interpretation must weaken. Keep compactness, separation, overlap, stability, uncertainty and practical usefulness on separate notebook lines. Explain once to a student and once to a parent. Add a counterexample involving a singleton, duplicated point, zero denominator, density bridge, unequal group size, local optimum or overlapping Gaussian components. State the extra evidence needed to distinguish explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.