Average Silhouette Width (ASW)

• Average Silhouette Width (ASW) is a clustering metric defined by averaging pointwise silhouette scores, which assess how well observations fit within their clusters compared to alternatives.
• It employs micro- and macro-averaging techniques to balance overall assignment quality with equitable cluster-wise evaluation, making it useful for both validation and optimization.
• ASW is widely applied in fields such as species delimitation, single-cell RNA-seq, and climate data analyses to estimate the ideal number of clusters.

Average Silhouette Width (ASW) is the global average of pointwise silhouette widths and is one of the most widely used internal criteria for assessing clustering quality and estimating the number of clusters. It summarizes, in a single scalar in $[-1,1]$, how strongly observations are associated with their assigned cluster relative to their nearest alternative cluster, thereby combining within-cluster compactness with between-cluster separation. In contemporary work, ASW appears both in its classical role as a validation index and as a direct optimization target for agglomerative, partitioning, and instance-weighted clustering procedures (Batool, 2019, Batool et al., 2019).

1. Pointwise silhouette and the ASW functional

For a dataset $X=\{x_1,\dots,x_n\}$, a clustering $\mathcal{C}_k=\{C_1,\dots,C_k\}$, a label function $l(i)\in\{1,\dots,k\}$, and cluster sizes $n_r=|C_r|$, the silhouette construction begins with two averages. The first is the mean intra-cluster distance

$a(i)=\frac{1}{n_{l(i)}-1}\sum_{\substack{l(h)=l(i)\h\neq i}} d(x_i,x_h),$

which measures how far $x_i$ is, on average, from the other points in its own cluster. The second is the mean nearest-cluster distance

$$b(i)=\min_{r\neq l(i)} \frac{1}{n_r}\sum_{l(h)=r} d(x_i,x_h),$$

which is the smallest average distance from $x_i$ to any other cluster. The silhouette width of observation $i$ is then

$X=\{x_1,\dots,x_n\}$0

and the Average Silhouette Width is

$X=\{x_1,\dots,x_n\}$1

The standard interpretation is geometric: $X=\{x_1,\dots,x_n\}$2 indicates that $X=\{x_1,\dots,x_n\}$3 is well embedded in its assigned cluster, $X=\{x_1,\dots,x_n\}$4 indicates a boundary case, and $X=\{x_1,\dots,x_n\}$5 suggests likely misassignment. ASW is the global average of these local diagnostics and is explicitly described as a global clustering quality measure that combines compactness and separation (Batool, 2019).

Most formulations assume a dissimilarity or distance satisfying non-negativity, zero diagonal, and symmetry. Some treatments work directly with a dissimilarity matrix and do not require the triangle inequality. Standard ASW is defined only for non-trivial partitions with at least two clusters; at the pointwise level, some later formulations set singleton silhouettes to $X=\{x_1,\dots,x_n\}$6, whereas ASW-driven hierarchical procedures instead avoid levels where $X=\{x_1,\dots,x_n\}$7 or $X=\{x_1,\dots,x_n\}$8 is undefined (Sträng et al., 10 Sep 2025).

2. Aggregation variants: micro, macro, and adaptive combinations

In standard usage, ASW almost always means the micro-averaged silhouette,

$X=\{x_1,\dots,x_n\}$9

which weights each observation equally. A less common alternative is the macro-averaged silhouette,

$\mathcal{C}_k=\{C_1,\dots,C_k\}$0

which first averages within clusters and then gives each cluster equal weight. The distinction becomes important under cluster-size imbalance: micro-averaging is more strongly influenced by larger clusters, whereas macro-averaging mitigates cluster-size bias by assigning equal weight to clusters irrespective of their sizes, but may overemphasize noise from under-represented groups (Semoglou et al., 15 Apr 2026).

This aggregation issue has become methodologically significant because many recent ASW-based algorithms allow the user to choose which aggregate to optimize. One example is K-Sil, which permits micro ASW, macro ASW, or a convex combination

$\mathcal{C}_k=\{C_1,\dots,C_k\}$1

In that framework, micro aggregation is associated with overall average assignment quality, macro aggregation with equitable cluster-wise quality, and hybrid aggregation with a trade-off between the two (Semoglou et al., 15 Jun 2025).

A further development is Composite Silhouette, which uses repeated subsampled clusterings, computes both micro- and macro-averaged silhouette scores on each subsample, and combines them through an adaptive convex weight driven by their normalized discrepancy and a bounded nonlinearity. The resulting criterion is then averaged across subsamples and used for cluster-count selection. This line of work treats ASW not as a single immutable scalar, but as the base quantity for a family of aggregation strategies that respond differently to imbalance and heterogeneity (Semoglou et al., 15 Apr 2026).

3. ASW as a clustering objective

A major methodological shift occurred when ASW was elevated from a post hoc validation index to an explicit clustering objective. In the OSil formulation, “Optimum Average Silhouette Width” clustering is defined as the partition maximizing ASW over all non-trivial labelings, and the corresponding algorithm performs greedy single-point reassignment: for fixed $\mathcal{C}_k=\{C_1,\dots,C_k\}$2, it evaluates all candidate moves of one observation to a different cluster, accepts the move with maximal ASW improvement, and stops at a local maximum under one-point changes. A fast subsampling-based approximation, FOSil, was introduced to reduce the computational burden on larger datasets (Batool, 2019, Batool et al., 2019).

A hierarchical version of the same idea is HOSil. Instead of using a classical linkage such as single, complete, average, or Ward, HOSil defines the linkage at each agglomerative step by examining all possible cluster pairs and merging the pair whose merger yields the highest ASW for the resulting partition. In this construction, ASW is the linkage objective as well as the criterion later used to cut the dendrogram, thereby unifying cluster construction and cluster-number estimation within one index (Batool, 2019).

ASW has also been incorporated into centroid-based refinement. K-Sil computes silhouettes at every iteration of a k-means-like procedure, converts those silhouettes into instance weights by either a power scheme or an exponential rank-based scheme, and updates centroids as weighted means. In that setting, silhouettes are not merely evaluative: they determine how much influence each point has on centroid movement, and the algorithm can target micro, macro, or hybrid ASW directly (Semoglou et al., 15 Jun 2025).

These developments share a common premise: if ASW is trusted as an internal definition of clustering quality, then it can be used not only to compare clusterings but also to generate them. The practical consequence is a family of clustering algorithms whose search dynamics are explicitly governed by compactness-separation trade-offs encoded in the silhouette formula.

4. Limitations, biases, and failure modes

A central limitation is that ASW is an internal, geometry-based criterion rather than a recovery criterion tied to a generating partition. Simulation studies repeatedly report that the clustering with the highest ASW need not coincide with the clustering with the highest adjusted Rand index, and that methods explicitly maximizing ASW can produce higher ASW than the true partition while recovering the data-generating labels less well. This establishes that maximizing ASW does not guarantee maximal agreement with ground truth (Batool, 2019, Batool, 2019).

A second limitation is a bias toward too few clusters, often $\mathcal{C}_k=\{C_1,\dots,C_k\}$3, when some subsets are very strongly separated but the remaining structure is more moderate. In such cases, ASW can prefer a partition that isolates a highly separated subset and lumps the rest together, because the increase in nearest-other-cluster distances can dominate the penalty from increased within-cluster distances. This behavior is emphasized both in simulation studies and in application examples involving outliers or coarse two-block decompositions (Batool et al., 2019).

A third limitation concerns diffuse, elongated, or otherwise wide clusters. Because ASW rewards small within-cluster dissimilarities, it can over-split broad structures into several compact subclusters. This behavior is documented for examples such as the “mouth” of the smiley dataset and the moon-like cluster in the Aggregation dataset, where ASW-driven procedures divide a semantically intended single cluster into smaller, more compact pieces (Batool, 2019).

A fourth limitation is difficulty with overlapping structures. One study states explicitly that HOSil “can not handle overlapping data structures,” and relates this directly to ASW’s preference for well-separated clusters. A plausible implication is that the same issue extends beyond a specific algorithm whenever ASW is the sole objective: overlapping but substantively meaningful groups may be replaced by more artificially separated partitions (Batool, 2019).

Finally, ASW is sensitive to the chosen dissimilarity. In one comparison using Euclidean, Manhattan, and Minkowski metrics, Minkowski produced the highest ASW values on average, Euclidean was intermediate, and Manhattan was lowest, with similar trends across clustering methods. This does not invalidate ASW, but it means that absolute ASW values and even the selected number of clusters are distance-dependent and must be interpreted relative to the underlying metric choice (Batool, 2019).

5. Computational and theoretical developments

Under general metrics, exact silhouette computation is inherently quadratic. For a clustering of $\mathcal{C}_k=\{C_1,\dots,C_k\}$4 elements into $\mathcal{C}_k=\{C_1,\dots,C_k\}$5 clusters, exact evaluation of all local silhouettes and the global ASW requires $\mathcal{C}_k=\{C_1,\dots,C_k\}$6 distance calculations, irrespective of $\mathcal{C}_k=\{C_1,\dots,C_k\}$7. This is the main obstacle to applying ASW routinely on massive datasets or repeatedly over many candidate values of $\mathcal{C}_k=\{C_1,\dots,C_k\}$8 (Sarpe et al., 2 Jul 2026).

Recent work has addressed this with rigorous approximation algorithms. A sampling-based framework estimates cluster-point distance sums and thereby approximates both local silhouettes and global ASW using

$\mathcal{C}_k=\{C_1,\dots,C_k\}$9

distance computations, while guaranteeing additive error $l(i)\in\{1,\dots,k\}$0 with probability at least $l(i)\in\{1,\dots,k\}$1. The same work gives scalable distributed implementations in MapReduce and MPC, with a constant number of rounds and sublinear local memory. These results are notable because earlier approximate methods were described as heuristics without provable and controllable guarantees (Sarpe et al., 2 Jul 2026).

A different theoretical advance concerns attainability. Although ASW is bounded above by $l(i)\in\{1,\dots,k\}$2, that upper limit is often not achievable for a fixed dataset. A recent analysis derives, for each observation, a sharp data-dependent upper bound on its silhouette width, then averages these bounds to obtain a canonical dataset-specific upper bound

$l(i)\in\{1,\dots,k\}$3

which often lies well below $l(i)\in\{1,\dots,k\}$4. The practical recommendation is to interpret an observed ASW relative to $l(i)\in\{1,\dots,k\}$5, not relative to the abstract ceiling $l(i)\in\{1,\dots,k\}$6; an ASW around $l(i)\in\{1,\dots,k\}$7 or $l(i)\in\{1,\dots,k\}$8 may already be near-optimal on a difficult dataset (Sträng et al., 10 Sep 2025).

ASW has also been examined axiomatically as a cluster quality measure. It has been shown to satisfy scale invariance, consistency, richness, and isomorphism invariance in the sense of the Ackerman–Ben-David framework for cluster quality functions. This theoretical result formalizes several intuitions already implicit in practice: rescaling all dissimilarities leaves ASW unchanged, shrinking within-cluster and enlarging between-cluster dissimilarities cannot reduce ASW, any non-trivial clustering can be made ASW-optimal for a suitable dissimilarity, and distance-preserving relabelings do not change the value (Batool et al., 2019).

6. Empirical role, applications, and current significance

ASW remains one of the dominant internal criteria for selecting the number of clusters. The classical rule is the maximum-ASW principle,

$l(i)\in\{1,\dots,k\}$9

and numerous studies still organize model selection around the ASW curve over $n_r=|C_r|$0. In the HOSil study, a large simulation comparing 11 indices across 10 data-generating processes reported 369 correct $n_r=|C_r|$1 estimates for HOSil, 356 for ASW with PAM, and 351 for PAMSIL; in the OSil initialization study, OASW via OSil achieved 844 correct estimates versus 809 for ASW alone across the compared combinations (Batool, 2019, Batool, 2019).

The empirical scope of ASW-based analysis is not restricted to Euclidean mixtures. Non-Euclidean applications include AFLP-based species delimitation with Jaccard distance and French rainfall extremes with the F-madogram. In the Veronica plant example, OSil and PAMSIL maximizing ASW over $n_r=|C_r|$2 recovered the expert-derived eight-species classification under Jaccard distance, whereas PAM with ASW selected seven clusters; in the rainfall example, ASW highlighted both a globally preferred $n_r=|C_r|$3 solution and scientifically more informative higher-$n_r=|C_r|$4 local maxima (Batool et al., 2019).

ASW-based methods have also been applied to single-cell RNA-seq, climate-region identification, and species delimitation. One HOSil example on the Goolam et al. mouse embryo data recovered the known $n_r=|C_r|$5 cell types with ASW $n_r=|C_r|$6 and ARI $n_r=|C_r|$7. Another example on Tetragonula bee data used shared-allele dissimilarity and found an ASW peak at $n_r=|C_r|$8, with ARI $n_r=|C_r|$9 relative to a nine-species morphological delimitation (Batool, 2019).

At the same time, more recent work increasingly treats plain micro-averaged ASW as a baseline rather than an endpoint. Composite Silhouette, for example, reports correct recovery of the true number of clusters on all 16 datasets in its experimental suite, specifically by aggregating micro- and macro-ASW over subsamples rather than relying on a single partition. This suggests not a displacement of ASW, but an expansion of its methodological role: ASW remains the core quantity, while current research refines how it is aggregated, optimized, approximated, and interpreted (Semoglou et al., 15 Apr 2026).

In that broader landscape, ASW occupies a dual position. It is simultaneously a diagnostic summary of cluster geometry and a design principle for clustering algorithms. Its enduring appeal lies in the directness of the underlying question it asks of each observation: is the point closer, on average, to its own cluster than to its best alternative? Its enduring difficulty lies in the fact that this question privileges isolated, compact partitions, which may or may not coincide with the substantive notion of a cluster in a given domain.