WHY MATHEMATICS?

Why Mathematics? | Adjusted Mutual Information for Clustering: Expected Agreement, Entropy Normalization, Chance Correction and Cluster Count

How observed and expected mutual information become a chance-adjusted clustering score, with a complete six-object calculation, variants and limits.

Why mathematics is important becomes easier to see when a real decision depends on it. This guide follows the mechanism from definitions and dimensions to a worked calculation, validation, practical learning and responsible use. It keeps factual evidence separate from interpretation and treats every illustrative scenario as an example rather than a promise of results.

Choose a reading route

Build mutual information from a contingency table
Calculate expected mutual information under fixed marginals
Apply entropy normalization after chance correction
Separate AMI from NMI, meaning and usefulness

Chapter map

1. Objects
2. Contingency table
3. Reference marginals
4. Proposed marginals
5. Reference entropy
6. Proposed entropy
7. Observed MI
8. Expected MI
9. Excess information
10. Arithmetic upper term
11. Adjusted denominator
12. AMI
13. Compare NMI
14. Label permutation
15. Negative values
16. Claim

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

1. Objects

A small example makes the mechanism visible. Use six aligned objects.

A concrete situation

Both partitions cover the same rows.

What the mathematics does not license

Join errors corrupt the contingency table. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Audit identifiers. 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. Contingency table

This is where modelling choices begin to affect evidence. Cross A,B with X,Y,Z.

A concrete situation

Counts are [[2,1,0],[0,1,2]].

What the mathematics does not license

Zeros contribute no observed MI term. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Show cells. 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. Reference marginals

The central mathematical idea is concrete. Sum rows.

A concrete situation

Counts are three and three.

What the mathematics does not license

The chance model conditions on marginals. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Record totals. 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. Proposed marginals

Here the geometry becomes operational. Sum columns.

A concrete situation

Counts are two, two and two.

What the mathematics does not license

Cluster count affects expected MI. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Record totals. 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.

Did You Know? In this chapter, the most important number is useful only together with its construction rule. Changing scaling, sample membership or validation design can change the numerical answer without changing the arithmetic.

5. Reference entropy

This step deserves a careful pause. Compute -2(.5 ln .5).

A concrete situation

H(C)=.693147 nats.

What the mathematics does not license

Name the log base. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Keep precision. 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. Proposed entropy

The useful question is not only how to calculate, but what the calculation assumes. Compute -3(1/3 ln 1/3).

A concrete situation

H(K)=1.098612 nats.

What the mathematics does not license

Entropy is not quality alone. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Keep precision. 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. Observed MI

A small example makes the mechanism visible. Sum cell probability times log ratio.

A concrete situation

MI=.462098 nats.

What the mathematics does not license

Raw MI rises with opportunities for matching. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Show 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.

8. Expected MI

This is where modelling choices begin to affect evidence. Average MI over fixed-marginal random tables.

A concrete situation

EMI=.277259 nats.

What the mathematics does not license

This is a combinatorial expectation. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

State model. 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. Excess information

The central mathematical idea is concrete. Subtract expected from observed.

A concrete situation

Numerator=.184839.

What the mathematics does not license

A negative numerator is possible. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Keep sign. 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. Arithmetic upper term

Here the geometry becomes operational. Average the two entropies.

A concrete situation

Mean entropy=.895880.

What the mathematics does not license

Other averaging rules give other variants. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Name method. 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. Adjusted denominator

This step deserves a careful pause. Subtract EMI from .895880.

A concrete situation

Denominator=.618621.

What the mathematics does not license

Do not divide by raw entropy mean. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Show subtraction. 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. AMI

The useful question is not only how to calculate, but what the calculation assumes. Divide .184839 by .618621.

A concrete situation

AMI=.298792.

What the mathematics does not license

Round only after division. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Cross-check software. 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. Compare NMI

A small example makes the mechanism visible. Calculate arithmetic NMI.

A concrete situation

NMI=.515804.

What the mathematics does not license

Normalization is not chance correction. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Explain gap. 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. Label permutation

This is where modelling choices begin to affect evidence. Rename labels bijectively.

A concrete situation

AMI is unchanged.

What the mathematics does not license

Objects must stay aligned. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Test invariance. 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. Negative values

The central mathematical idea is concrete. Consider less agreement than expected.

A concrete situation

AMI may be negative.

What the mathematics does not license

Zero is model-relative, not absence of structure. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Interpret baseline. 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

Here the geometry becomes operational. Name marginals, model and averaging method.

A concrete situation

Teaching AMI=.298792.

What the mathematics does not license

Do not claim semantic truth. Evidence becomes stronger when the limitation is tested instead of hidden.

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

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

Adjusted mutual information compares two partitions through the information their labels share, then removes the mutual information expected from random partitions with the observed cluster-size marginals. The score answers a sharper question than raw mutual information or normalized mutual information: how much agreement remains beyond a stated chance baseline when the number and sizes of groups can themselves create apparent information?

Use six aligned objects with reference labels A,A,A,B,B,B and proposed labels X,X,Y,Y,Z,Z. Their contingency table has non-zero counts 2,1,1,2. Mutual information is .462098 nats, the reference entropy is ln2=.693147 and the proposed entropy is ln3=1.098612. Under the fixed-marginal permutation model, expected mutual information is .277259 nats.

With the arithmetic-mean normalizer, the upper normalization term is (.693147+1.098612)/2=.895880. AMI=[.462098-.277259]/[.895880-.277259]=.298792. The corresponding arithmetic NMI is .515804, so chance correction materially changes the story for this small example. The expected term is a combinatorial expectation, not a number to estimate casually from a few random shuffles.

AMI equals one for identical partitions up to label renaming, is near zero on average for independent random partitions under its model and may be negative. Its value depends on the averaging method used in the normalizer and on the chance model. It does not establish semantic meaning, stable clusters, causal structure or fair decisions. Report the contingency table, marginals, entropy convention, expected MI, averaging rule, cluster count and uncertainty.

Frequently asked questions

What is adjusted mutual information?

Mutual information between two partitions minus its fixed-marginal chance expectation, divided by the corresponding adjusted normalization range.

Why adjust for chance?

Random partitions can share positive mutual information, especially with small samples or many clusters.

What is the expected MI?

The mean mutual information over random contingency tables with the observed row and column marginals fixed.

Can AMI be negative?

Yes. Agreement can fall below the stated chance expectation.

Why name the averaging method?

Minimum, geometric, arithmetic and maximum entropy normalizers can produce different AMI variants.

What must be reported?

Objects, table, marginals, entropies, observed MI, expected MI, averaging method, cluster count, software convention and uncertainty.

Primary source and fact-check note

Nguyen Xuan Vinh, Julien Epps and James Bailey, “Information Theoretic Measures for Clusterings Comparison: Variants, Properties, Normalization and Correction for Chance,” JMLR 11, 2837–2854 (2010). 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

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: clustering evidence and responsible use

Object audit

Freeze evaluated objects, exclusions, missing-value rules, duplicates and the join between reference and proposed labels.

Contingency audit

Show every cell, row total, column total, class size, cluster size and any tie before compressing evidence into a score.

Formula audit

Show observed and expected information, conditional entropies, majority cells, item intersections and every denominator.

Convention audit

Name logarithm base, entropy normalizer, chance model, self-inclusion, averaging weights, label model and edge-case rules.

Did You Know? A sophisticated adjustment can be algebraically correct and still answer the wrong scientific question when samples, labels, batches or validation boundaries are wrong.

Degeneracy audit

Test singleton clusters, one giant cluster, zero entropy, empty cells, class imbalance and denominator limits.

Stability audit

Repeat across seeds, bootstrap samples, perturbations, label uncertainty and plausible numbers of clusters.

Meaning audit

Compare numeric agreement with domain evidence without assuming that reference classes are complete, causal or fair.

Final claim

State exact partition agreement or distance evidence and stop before claims of truth, causation or guaranteed usefulness.

Did You Know? A sophisticated adjustment can be algebraically correct and still answer the wrong scientific question when samples, labels, batches or validation boundaries are wrong.

Mechanism workshop: calculate, challenge and transfer

Workshop 1: Objects

Begin with this clustering mechanism: Use six aligned objects. Recreate the concrete case: Both partitions cover the same rows. Write object IDs, both label vectors, contingency cells, marginals, entropies, observed MI, expected MI, averaging method and AMI with identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Join errors corrupt the contingency table. Complete the transfer step: Audit identifiers. Change exactly one object, cluster, class, contingency cell, marginal, cluster count, logarithm base, expected-information model, normalizer, winning cell, tie, per-object intersection, averaging weight or label uncertainty. Explain which mathematical object changes, which score moves and which interpretation must weaken. Keep raw agreement, chance correction, directional error, metric distance, stability and practical usefulness on separate notebook lines. Explain once to a student and once to a parent. Add a counterexample involving singleton clusters, one giant cluster, severe imbalance, renamed labels, overlapping membership or an uncertain reference. State the extra evidence needed to distinguish explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 2: Contingency table

Begin with this clustering mechanism: Cross A,B with X,Y,Z. Recreate the concrete case: Counts are [[2,1,0],[0,1,2]]. Write object IDs, both label vectors, contingency cells, marginals, entropies, observed MI, expected MI, averaging method and AMI with identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Zeros contribute no observed MI term. Complete the transfer step: Show cells. Change exactly one object, cluster, class, contingency cell, marginal, cluster count, logarithm base, expected-information model, normalizer, winning cell, tie, per-object intersection, averaging weight or label uncertainty. Explain which mathematical object changes, which score moves and which interpretation must weaken. Keep raw agreement, chance correction, directional error, metric distance, stability and practical usefulness on separate notebook lines. Explain once to a student and once to a parent. Add a counterexample involving singleton clusters, one giant cluster, severe imbalance, renamed labels, overlapping membership or an uncertain reference. State the extra evidence needed to distinguish explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 3: Reference marginals

Begin with this clustering mechanism: Sum rows. Recreate the concrete case: Counts are three and three. Write object IDs, both label vectors, contingency cells, marginals, entropies, observed MI, expected MI, averaging method and AMI with identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: The chance model conditions on marginals. Complete the transfer step: Record totals. Change exactly one object, cluster, class, contingency cell, marginal, cluster count, logarithm base, expected-information model, normalizer, winning cell, tie, per-object intersection, averaging weight or label uncertainty. Explain which mathematical object changes, which score moves and which interpretation must weaken. Keep raw agreement, chance correction, directional error, metric distance, stability and practical usefulness on separate notebook lines. Explain once to a student and once to a parent. Add a counterexample involving singleton clusters, one giant cluster, severe imbalance, renamed labels, overlapping membership or an uncertain reference. State the extra evidence needed to distinguish explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 4: Proposed marginals

Begin with this clustering mechanism: Sum columns. Recreate the concrete case: Counts are two, two and two. Write object IDs, both label vectors, contingency cells, marginals, entropies, observed MI, expected MI, averaging method and AMI with identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Cluster count affects expected MI. Complete the transfer step: Record totals. Change exactly one object, cluster, class, contingency cell, marginal, cluster count, logarithm base, expected-information model, normalizer, winning cell, tie, per-object intersection, averaging weight or label uncertainty. Explain which mathematical object changes, which score moves and which interpretation must weaken. Keep raw agreement, chance correction, directional error, metric distance, stability and practical usefulness on separate notebook lines. Explain once to a student and once to a parent. Add a counterexample involving singleton clusters, one giant cluster, severe imbalance, renamed labels, overlapping membership or an uncertain reference. State the extra evidence needed to distinguish explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Did You Know? A sophisticated adjustment can be algebraically correct and still answer the wrong scientific question when samples, labels, batches or validation boundaries are wrong.

Workshop 5: Reference entropy

Begin with this clustering mechanism: Compute -2(.5 ln .5). Recreate the concrete case: H(C)=.693147 nats. Write object IDs, both label vectors, contingency cells, marginals, entropies, observed MI, expected MI, averaging method and AMI with identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Name the log base. Complete the transfer step: Keep precision. Change exactly one object, cluster, class, contingency cell, marginal, cluster count, logarithm base, expected-information model, normalizer, winning cell, tie, per-object intersection, averaging weight or label uncertainty. Explain which mathematical object changes, which score moves and which interpretation must weaken. Keep raw agreement, chance correction, directional error, metric distance, stability and practical usefulness on separate notebook lines. Explain once to a student and once to a parent. Add a counterexample involving singleton clusters, one giant cluster, severe imbalance, renamed labels, overlapping membership or an uncertain reference. State the extra evidence needed to distinguish explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 6: Proposed entropy

Begin with this clustering mechanism: Compute -3(1/3 ln 1/3). Recreate the concrete case: H(K)=1.098612 nats. Write object IDs, both label vectors, contingency cells, marginals, entropies, observed MI, expected MI, averaging method and AMI with identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Entropy is not quality alone. Complete the transfer step: Keep precision. Change exactly one object, cluster, class, contingency cell, marginal, cluster count, logarithm base, expected-information model, normalizer, winning cell, tie, per-object intersection, averaging weight or label uncertainty. Explain which mathematical object changes, which score moves and which interpretation must weaken. Keep raw agreement, chance correction, directional error, metric distance, stability and practical usefulness on separate notebook lines. Explain once to a student and once to a parent. Add a counterexample involving singleton clusters, one giant cluster, severe imbalance, renamed labels, overlapping membership or an uncertain reference. State the extra evidence needed to distinguish explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 7: Observed MI

Begin with this clustering mechanism: Sum cell probability times log ratio. Recreate the concrete case: MI=.462098 nats. Write object IDs, both label vectors, contingency cells, marginals, entropies, observed MI, expected MI, averaging method and AMI with identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Raw MI rises with opportunities for matching. Complete the transfer step: Show formula. Change exactly one object, cluster, class, contingency cell, marginal, cluster count, logarithm base, expected-information model, normalizer, winning cell, tie, per-object intersection, averaging weight or label uncertainty. Explain which mathematical object changes, which score moves and which interpretation must weaken. Keep raw agreement, chance correction, directional error, metric distance, stability and practical usefulness on separate notebook lines. Explain once to a student and once to a parent. Add a counterexample involving singleton clusters, one giant cluster, severe imbalance, renamed labels, overlapping membership or an uncertain reference. State the extra evidence needed to distinguish explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 8: Expected MI

Begin with this clustering mechanism: Average MI over fixed-marginal random tables. Recreate the concrete case: EMI=.277259 nats. Write object IDs, both label vectors, contingency cells, marginals, entropies, observed MI, expected MI, averaging method and AMI with identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: This is a combinatorial expectation. Complete the transfer step: State model. Change exactly one object, cluster, class, contingency cell, marginal, cluster count, logarithm base, expected-information model, normalizer, winning cell, tie, per-object intersection, averaging weight or label uncertainty. Explain which mathematical object changes, which score moves and which interpretation must weaken. Keep raw agreement, chance correction, directional error, metric distance, stability and practical usefulness on separate notebook lines. Explain once to a student and once to a parent. Add a counterexample involving singleton clusters, one giant cluster, severe imbalance, renamed labels, overlapping membership or an uncertain reference. State the extra evidence needed to distinguish explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Did You Know? A sophisticated adjustment can be algebraically correct and still answer the wrong scientific question when samples, labels, batches or validation boundaries are wrong.

Workshop 9: Excess information

Begin with this clustering mechanism: Subtract expected from observed. Recreate the concrete case: Numerator=.184839. Write object IDs, both label vectors, contingency cells, marginals, entropies, observed MI, expected MI, averaging method and AMI with identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: A negative numerator is possible. Complete the transfer step: Keep sign. Change exactly one object, cluster, class, contingency cell, marginal, cluster count, logarithm base, expected-information model, normalizer, winning cell, tie, per-object intersection, averaging weight or label uncertainty. Explain which mathematical object changes, which score moves and which interpretation must weaken. Keep raw agreement, chance correction, directional error, metric distance, stability and practical usefulness on separate notebook lines. Explain once to a student and once to a parent. Add a counterexample involving singleton clusters, one giant cluster, severe imbalance, renamed labels, overlapping membership or an uncertain reference. State the extra evidence needed to distinguish explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 10: Arithmetic upper term

Begin with this clustering mechanism: Average the two entropies. Recreate the concrete case: Mean entropy=.895880. Write object IDs, both label vectors, contingency cells, marginals, entropies, observed MI, expected MI, averaging method and AMI with identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Other averaging rules give other variants. Complete the transfer step: Name method. Change exactly one object, cluster, class, contingency cell, marginal, cluster count, logarithm base, expected-information model, normalizer, winning cell, tie, per-object intersection, averaging weight or label uncertainty. Explain which mathematical object changes, which score moves and which interpretation must weaken. Keep raw agreement, chance correction, directional error, metric distance, stability and practical usefulness on separate notebook lines. Explain once to a student and once to a parent. Add a counterexample involving singleton clusters, one giant cluster, severe imbalance, renamed labels, overlapping membership or an uncertain reference. State the extra evidence needed to distinguish explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 11: Adjusted denominator

Begin with this clustering mechanism: Subtract EMI from .895880. Recreate the concrete case: Denominator=.618621. Write object IDs, both label vectors, contingency cells, marginals, entropies, observed MI, expected MI, averaging method and AMI with identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Do not divide by raw entropy mean. Complete the transfer step: Show subtraction. Change exactly one object, cluster, class, contingency cell, marginal, cluster count, logarithm base, expected-information model, normalizer, winning cell, tie, per-object intersection, averaging weight or label uncertainty. Explain which mathematical object changes, which score moves and which interpretation must weaken. Keep raw agreement, chance correction, directional error, metric distance, stability and practical usefulness on separate notebook lines. Explain once to a student and once to a parent. Add a counterexample involving singleton clusters, one giant cluster, severe imbalance, renamed labels, overlapping membership or an uncertain reference. State the extra evidence needed to distinguish explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 12: AMI

Begin with this clustering mechanism: Divide .184839 by .618621. Recreate the concrete case: AMI=.298792. Write object IDs, both label vectors, contingency cells, marginals, entropies, observed MI, expected MI, averaging method and AMI with identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Round only after division. Complete the transfer step: Cross-check software. Change exactly one object, cluster, class, contingency cell, marginal, cluster count, logarithm base, expected-information model, normalizer, winning cell, tie, per-object intersection, averaging weight or label uncertainty. Explain which mathematical object changes, which score moves and which interpretation must weaken. Keep raw agreement, chance correction, directional error, metric distance, stability and practical usefulness on separate notebook lines. Explain once to a student and once to a parent. Add a counterexample involving singleton clusters, one giant cluster, severe imbalance, renamed labels, overlapping membership or an uncertain reference. State the extra evidence needed to distinguish explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Did You Know? A sophisticated adjustment can be algebraically correct and still answer the wrong scientific question when samples, labels, batches or validation boundaries are wrong.

Workshop 13: Compare NMI

Begin with this clustering mechanism: Calculate arithmetic NMI. Recreate the concrete case: NMI=.515804. Write object IDs, both label vectors, contingency cells, marginals, entropies, observed MI, expected MI, averaging method and AMI with identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Normalization is not chance correction. Complete the transfer step: Explain gap. Change exactly one object, cluster, class, contingency cell, marginal, cluster count, logarithm base, expected-information model, normalizer, winning cell, tie, per-object intersection, averaging weight or label uncertainty. Explain which mathematical object changes, which score moves and which interpretation must weaken. Keep raw agreement, chance correction, directional error, metric distance, stability and practical usefulness on separate notebook lines. Explain once to a student and once to a parent. Add a counterexample involving singleton clusters, one giant cluster, severe imbalance, renamed labels, overlapping membership or an uncertain reference. State the extra evidence needed to distinguish explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 14: Label permutation

Begin with this clustering mechanism: Rename labels bijectively. Recreate the concrete case: AMI is unchanged. Write object IDs, both label vectors, contingency cells, marginals, entropies, observed MI, expected MI, averaging method and AMI with identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Objects must stay aligned. Complete the transfer step: Test invariance. Change exactly one object, cluster, class, contingency cell, marginal, cluster count, logarithm base, expected-information model, normalizer, winning cell, tie, per-object intersection, averaging weight or label uncertainty. Explain which mathematical object changes, which score moves and which interpretation must weaken. Keep raw agreement, chance correction, directional error, metric distance, stability and practical usefulness on separate notebook lines. Explain once to a student and once to a parent. Add a counterexample involving singleton clusters, one giant cluster, severe imbalance, renamed labels, overlapping membership or an uncertain reference. State the extra evidence needed to distinguish explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 15: Negative values

Begin with this clustering mechanism: Consider less agreement than expected. Recreate the concrete case: AMI may be negative. Write object IDs, both label vectors, contingency cells, marginals, entropies, observed MI, expected MI, averaging method and AMI with identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Zero is model-relative, not absence of structure. Complete the transfer step: Interpret baseline. Change exactly one object, cluster, class, contingency cell, marginal, cluster count, logarithm base, expected-information model, normalizer, winning cell, tie, per-object intersection, averaging weight or label uncertainty. Explain which mathematical object changes, which score moves and which interpretation must weaken. Keep raw agreement, chance correction, directional error, metric distance, stability and practical usefulness on separate notebook lines. Explain once to a student and once to a parent. Add a counterexample involving singleton clusters, one giant cluster, severe imbalance, renamed labels, overlapping membership or an uncertain reference. 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 clustering mechanism: Name marginals, model and averaging method. Recreate the concrete case: Teaching AMI=.298792. Write object IDs, both label vectors, contingency cells, marginals, entropies, observed MI, expected MI, averaging method and AMI with identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Do not claim semantic truth. Complete the transfer step: Bound conclusion. Change exactly one object, cluster, class, contingency cell, marginal, cluster count, logarithm base, expected-information model, normalizer, winning cell, tie, per-object intersection, averaging weight or label uncertainty. Explain which mathematical object changes, which score moves and which interpretation must weaken. Keep raw agreement, chance correction, directional error, metric distance, stability and practical usefulness on separate notebook lines. Explain once to a student and once to a parent. Add a counterexample involving singleton clusters, one giant cluster, severe imbalance, renamed labels, overlapping membership or an uncertain reference. State the extra evidence needed to distinguish explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.