Fowlkes–Mallows Index¶

Compare two partitions by the geometric mean of pairwise co-membership precision and recall, rewarding pairs clustered together by both while excluding true-negative pairs from the score.

Version: v3 · 2026-09-19 · History
Domain-specific #: 1874
Domain group: Formal Sciences
Origin domain: Experimental Design & Statistics
Subdomain: Cluster Analysis → Experimental Design & Statistics
Aliases: Fowlkes–Mallows measure, FM index, Fowlkes–Mallows score, B K statistic

Core Idea¶

The Fowlkes–Mallows Index compares two partitions of the same finite item set through unordered item pairs. Let \(TP\) count pairs placed in the same block by both partitions, \(FP\) pairs co-clustered only by the first, and \(FN\) pairs co-clustered only by the second. When both denominators are positive, \(\mathrm{FM}=TP/\sqrt{(TP+FP)(TP+FN)}\). Equivalently, it is \(\sqrt{PR}\), the geometric mean of pairwise precision \(P=TP/(TP+FP)\) and pairwise recall \(R=TP/(TP+FN)\). The score lies in \([0,1]\); one means the two partitions induce identical pairwise co-membership relations.

How would you explain it like I'm…

Same-Box Pairs Score

Two kids each sort the same pile of toys into groups. To see how much they agree, you look at every pair of toys and ask: did both kids put these two together? The Fowlkes–Mallows Index turns that into a score from 0 to 1, where 1 means they grouped every pair the same way. Pairs that both kids kept apart don't add to the score.

Pair-Matching Sorting Score

The Fowlkes–Mallows Index compares two ways of sorting the same things into groups. It looks at every pair of items and counts: pairs both sortings put together, pairs only the first put together, and pairs only the second put together. From these it makes a score between 0 and 1, where 1 means the groupings match perfectly. It ignores pairs that both sortings kept apart, which can be a huge number. That's useful, but it also means the score has quirks, like changing when there are lots of tiny or uneven groups.

Pairwise Precision-Recall Cluster Agreement

The Fowlkes–Mallows Index compares two clusterings, or partitions, of the same items by looking at item pairs. TP is the number of pairs grouped together by both, FP the number grouped together only by the first, and FN the number grouped together only by the second. Precision is TP/(TP+FP), recall is TP/(TP+FN), and the index is the square root of their product, their geometric mean. It ranges from 0 to 1, with 1 meaning identical pairwise grouping. Unlike the plain Rand index, it ignores pairs separated by both clusterings, so the huge number of such pairs cannot inflate the score. It has no correction for agreement that would happen by chance, and it can shift with cluster sizes. It is a similarity, not a distance, and it can be undefined if, for example, one partition puts no pairs together.

The Fowlkes–Mallows Index compares two partitions of a finite set via unordered item pairs: TP = pairs co-clustered in both, FP = co-clustered only in the first, FN = co-clustered only in the second. FM = TP/√((TP+FP)(TP+FN)) = √(P·R), the geometric mean of pairwise precision P and recall R, defined when both denominators are positive. It lies in [0,1], equaling 1 exactly when the partitions induce identical co-membership relations. Fowlkes and Mallows introduced it for hierarchical clusterings, cutting both trees into k clusters and computing B_k from contingency-table sums of squares (counts that are twice the pair counts, so the factor cancels), then plotting B_k against k to view agreement across levels. Ignoring true negatives keeps it from being dominated by jointly separated pairs, unlike the unadjusted Rand index, but it lacks chance correction, is sensitive to cluster-size distribution, and can shift with many singletons or imbalanced blocks. It is symmetric in the two partitions, yet it is a similarity index, not a metric; 1−FM need not satisfy the triangle inequality. Degenerate partitions can zero a denominator, so implementations must state how that case is handled.

Scope of Application¶

The Fowlkes–Mallows Index is literal when two hard partitions of the same finite objects are compared by agreement on which pairs belong together.

External cluster validation. An estimated partition is compared with a benchmark classification.
Method comparison. Outputs from different clustering algorithms are compared without matching arbitrary labels.
Stability analysis. Bootstrap or perturbation runs are compared pairwise to assess co-membership stability.
Hierarchical clustering. Dendrogram cuts are compared over a range of cluster counts using \(B_k\) curves.
Consensus clustering. Candidate consensus partitions are scored against contributing partitions.
Image segmentation. Region partitions of a common pixel or object universe are externally compared.
Bioinformatics. Sample, cell, gene, or sequence partitions are compared under a fixed item universe.
Benchmark reporting. FM complements indices with different true-negative and chance-baseline behavior.

Clarity¶

Declare the common item universe, whether pairs are unordered and exclude self-pairs, which partition supplies the precision denominator, the hard-partition assumption, and the zero-denominator convention. Report \(TP\), \(FP\), and \(FN\) or the contingency table alongside the score so it can be reproduced. For hierarchies, name every cut rule and cluster count. Do not call FM a metric without defining and proving a separate distance. Do not interpret a raw value against a universal chance baseline; the expected value depends on the cluster-size and random-partition model.

Manages Complexity¶

Comparing partitions naively requires matching arbitrary block labels and reasoning over whole set systems. FM removes label identity by reducing both partitions to the same binary relation on item pairs, then compresses three relevant counts into a symmetric scalar. The geometric mean balances one-sided over-merging and over-splitting: high precision with low recall, or the reverse, cannot yield a high final score. This compression hides structure.

Abstract Reasoning¶

Verify that both inputs partition the same item universe.
Enumerate or derive counts for every unordered pair's two co-membership decisions.
Compute \(TP\), \(FP\), and \(FN\) without using arbitrary cluster labels.
Check whether \(TP+FP\) and \(TP+FN\) are positive.
Apply and disclose the chosen convention if either denominator vanishes.
Compute pairwise precision and recall, then their geometric mean. 7.

Knowledge Transfer¶

The strict parent is Comparison. FM takes two structured outputs on a common universe and produces a rule-governed similarity judgment from their relational agreements and disagreements. The parent transfers the need for aligned objects, declared criteria, and interpretable direction. The domain residual is pairwise co-membership, exclusion of true negatives, and geometric aggregation of precision and recall. Clustering is the object domain but cannot be a specialization parent because FM compares clusterings rather than constructing one.

Relationships to Other Abstractions¶

Current abstraction Fowlkes–Mallows Index Domain-specific

Parents (1) — outgoing typed relations

Fowlkes–Mallows Index is a kind of Comparison Prime

Comparison is the strict parent by composition/presupposition: FM compares two partitions under a fixed relational criterion and returns an ordered similarity judgment.

Hierarchy path (1) — routes to 1 parentless root

Fowlkes–Mallows Index → is a kind of Comparison

Neighborhood in Abstraction Space¶

Fowlkes–Mallows Index sits in a sparse region of the domain-specific corpus (97^th percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.

Family — Unclustered & Miscellaneous (2769 abstractions)

Nearest neighbors

Computed from structural-signature embeddings · 2026-10-25