Visual Assessment of Cluster Tendency (VAT)

Updated 7 July 2026

VAT is a visual technique that reorders pairwise dissimilarity matrices to produce grayscale images where dark blocks indicate potential clusters.
It employs MST-based, greedy reordering methods, linking to single-linkage clustering and inspiring variants like iVAT, SpecVAT, ConiVAT, and DeepVAT.
VAT serves as a diagnostic tool for cluster tendency assessment and helps identify misalignments in high-dimensional data representations across audio, image, and incremental clustering applications.

Visual Assessment of Cluster Tendency (VAT) is a visual technique for determining the potential cluster structure and the possible number of clusters in numerical data. In its standard form, it is a purely unsupervised, matrix-reordering method that takes a full pairwise dissimilarity matrix and displays a reordered version as a grayscale image, where contiguous dark blocks near the main diagonal indicate groups of mutually similar objects. Within the literature represented here, VAT is treated both as a cluster-tendency diagnostic and as a companion to single-linkage clustering, and it has generated a family of extensions including iVAT, SpecVAT, ConiVAT, and deep-embedding-based pipelines for audio and image data (Rathore et al., 2020, Heise et al., 2021).

1. Formal definition and image geometry

Let $X=\{x_1,\dots,x_n\}\subset\mathbb{R}^p$ . VAT begins from an $n\times n$ dissimilarity matrix $D=[d_{ij}]$ , where a common choice is Euclidean distance,

$d(i,j)=\|x_i-x_j\|_2,$

although Manhattan distance and cosine dissimilarity are also explicitly discussed in later implementations. Once a permutation $P$ of the indices is computed, the reordered matrix is written as $R=D(P,P)$ or $D^\ast_{ij}=D_{\pi[i],\pi[j]}$ , depending on notation. Rendering this reordered matrix as a grayscale image produces the VAT image, in which dark pixels correspond to small pairwise dissimilarities and light pixels correspond to large ones (Siddique et al., 2018, Avinash et al., 21 Jul 2025).

The semantics of the image are central to the method. The main diagonal is zero-valued, since each object is identical to itself, and therefore appears black. If clusters are present, contiguous off-diagonal regions of low dissimilarity appear as dark diagonal blocks. Bright regions between blocks indicate separation between groups. In the practical reading of VAT plots, the number of dark blocks is used as an estimate of the number of clusters, block size reflects cluster size, and sharp bright separations indicate well-separated groups (Rathore et al., 2020, Siddique et al., 2018).

A normalization step is often applied before display. One formulation uses $I=D^\ast/\max(D^\ast)$ , exploiting the zero diagonal. The visual content, however, resides in the relative contrast of intra-cluster versus inter-cluster dissimilarities rather than in any single normalization convention (Siddique et al., 2018).

2. Reordering mechanisms and relation to single linkage

VAT is consistently described as an MST-based or Prim-like reordering, but the supplied sources present more than one constructive form. In one exposition, VAT grows a minimum-spanning tree over the complete graph with edge weights $D$ , using a Prim-like rule in which the next object is the one with smallest distance to the set of already selected objects; the resulting permutation is written $P=\mathrm{VAT}(D)$ (Rathore et al., 2020). Other expositions initialize from two highly separated points, or from a point selected by a row-sum heuristic and then its farthest counterpart, and then iteratively choose the unselected index maximizing its minimum distance to the selected set, i.e. a maximin or farthest-first procedure (Avinash et al., 21 Jul 2025, Silva et al., 2018).

This suggests that the literature treats the reordered dissimilarity image as the invariant object of interest, while allowing implementation-level variation in the exact initialization and update convention. In all cases, the procedure is greedy, uses the full dissimilarity matrix, and attempts to place mutually similar objects in contiguous positions so that block structure becomes visually salient (Avinash et al., 21 Jul 2025, Siddique et al., 2018).

VAT is also explicitly tied to single-linkage (SL) hierarchical clustering. A back-pass through the same MST yields SL clusters, and cutting the $n\times n$ 0 largest MST edges splits the data into $n\times n$ 1 SL-clusters that align with the dark diagonal blocks in the VAT image. This relation is exploited directly in ConiVAT, and it also motivates the use of VAT as a preprocessing order for incremental ART-based clustering, where the VAT sequence approximates a single-linkage MST traversal and thereby reduces order dependence (Rathore et al., 2020, Silva et al., 2018).

3. Reading VAT images and estimating cluster number

The classic appeal of VAT is that it requires no parameters beyond the dissimilarities themselves; the cluster number $n\times n$ 2 can be chosen by “eyeballing” the number of dark blocks. In the simplest two-cluster scenarios, the reordered matrix exhibits two contiguous dark squares along the diagonal and a bright gap between them. More generally, gradual transitions or fuzzy blocks suggest overlapping or non-compact structure, isolated dark pixels off the diagonal may indicate outliers, and long thin blocks or chains may signal chain-shaped or non-convex structures that VAT does not clearly resolve (Rathore et al., 2020, Siddique et al., 2018).

For applications requiring automated cluster-count estimation, the literature also reports explicit post-processing. In a large-scale environmental-audio study, the authors used Sledge et al.’s Cluster-Count Extraction (CCE) algorithm. Their procedure thresholds the VAT image using Otsu’s method, builds a histogram of an off-diagonal slice of the reordered matrix, and locates peaks above a parameter $n\times n$ 3, where $n\times n$ 4 was set to half the maximum histogram value rather than zero. In that study, CCE counted 42 clusters in the full $n\times n$ 5 VAT image, substantially exceeding the six city labels and ten scene labels used as annotations, which the authors interpreted as evidence of latent multi-label or subcluster structure (Heise et al., 2021).

The interpretive status of a VAT image is therefore dual. It is a qualitative visualization intended to reveal cluster tendency without fixing $n\times n$ 6 a priori, but it can also support quantitative cluster-count extraction and downstream clustering decisions when combined with explicit thresholding or MST cuts (Heise et al., 2021, Rathore et al., 2020).

4. Variants in the VAT family

Several variants modify either the dissimilarity itself or the representation on which VAT operates.

Variant	Key mechanism	Stated purpose
iVAT	Path-based minimax distance transform	Improve VAT for “tough” data
SpecVAT	Spectral embedding from locally weighted affinity	Visually “denoised” block structure
ConiVAT	Constraints, metric learning, minimum-transitive-dissimilarity	Improve VAT/iVAT on challenging and complex datasets
DeepVAT	Self-supervised embeddings, t-SNE, MMRS, iVAT	Assess cluster structure in image datasets

The improved VAT, or iVAT, replaces each direct dissimilarity $n\times n$ 7 by a path-based minimax distance

$n\times n$ 8

equivalently the all-pairs minimax or bottleneck transform of $n\times n$ 9. The stated intuition is that two points in the same dense region may be connected by a path whose maximum inter-point gap is small even when their direct dissimilarity is large. After computing $D=[d_{ij}]$ 0, VAT is run on $D=[d_{ij}]$ 1, producing sharper and more uniform diagonal blocks when direct distances are misleading (Rathore et al., 2020).

SpecVAT applies the same visual logic in a spectral space. It first forms a locally weighted affinity

$D=[d_{ij}]$ 2

constructs the degree matrix $D=[d_{ij}]$ 3, computes the normalized graph Laplacian $D=[d_{ij}]$ 4, and uses the rows of the $D=[d_{ij}]$ 5 smallest nonzero eigenvectors as an embedding before applying VAT. In the cited audio study, $D=[d_{ij}]$ 6 was selected via A-SpecVAT, and the resulting SpecVAT images appeared visually “denoised,” with sharper dark blocks and fewer off-diagonal artefacts (Heise et al., 2021).

ConiVAT extends iVAT to the semi-supervised setting by incorporating pairwise must-link and cannot-link constraints. Its pipeline optionally expands constraints by transitive closure, learns a Mahalanobis metric

$D=[d_{ij}]$ 7

using the Xing et al. convex formulation, forces must-link pairs to zero, computes a minimum-transitive-dissimilarity transform exactly as in iVAT, and then runs VAT on the constrained matrix. The paper reports that this improves the quality of iVAT images for complex datasets and overcomes limitations of SL clustering with VAT/iVAT due to “noisy” bridges between clusters (Rathore et al., 2020).

DeepVAT addresses image data, where raw-pixel dissimilarities are often uninformative. It uses a self-supervised deep neural network, specifically a SimCLR-style encoder and projection head, to generate representative embeddings; these embeddings are reduced to two dimensions using t-SNE and then passed to VAT-based algorithms. For very large datasets, it further introduces MMRS, a smart subsampling strategy based on maximin prototypes and group-wise random sampling (Mazumder et al., 2023).

5. Use in audio, image, and incremental clustering studies

In environmental audio, VAT and SpecVAT were used to analyze the DCASE 2018 ASC Task 1A dataset, comprising ten acoustic scenes, six cities, and $D=[d_{ij}]$ 8 ten-second binaural recordings. Each recording was represented by a 128-dimensional vector formed by taking the feature-wise mean over 431 frames of a 128-band log-mel spectrogram computed with a 2048-point STFT, hop length 512, and sampling rate $D=[d_{ij}]$ 9 Hz. On the full dataset, VAT combined with CCE produced 42 clusters, and zoomed-in windows showed dark blocks corresponding cleanly to scenes such as street_traffic, public_square, or bus. The same study also reported that VAT and SpecVAT corroborated supervised classifier confusions: label pairs frequently confused in prior work, such as airport versus shopping_mall and airport versus public_square versus street_pedestrian, appeared as interleaved or weakly separated VAT blocks, whereas park and street_traffic were visually distinct and easier to classify (Heise et al., 2021).

VAT has also been used as a preprocessing stage for incremental clustering systems. In the distributed dual vigilance fuzzy adaptive resonance theory framework, VAT ordering is applied before one-pass DDVFA learning in offline mode. On 30 benchmark datasets, VAT+DDVFA improved mean Adjusted Rand Index from approximately $P$ 6 under random-order DDVFA to approximately $P$ 7, with a statistically significant improvement, and on the Spiral dataset the reported AR improved from about $P$ 8 to $P$ 9. The authors found VAT+DDVFA statistically equivalent to DDVFA cascaded with Merge ART in offline mode, while preserving a batch-mode remedy for order sensitivity (Silva et al., 2018).

6. Complexity, scaling, and optimized implementations

The computational burden of VAT comes from two sources: the full pairwise dissimilarity matrix and the reordering itself. One implementation analysis reports $R=D(P,P)$ 0 time to compute all pairwise distances, $R=D(P,P)$ 1 time for the reordering step, and $R=D(P,P)$ 2 memory to store $R=D(P,P)$ 3. Another source summarizes the standard implementation as suffering from $R=D(P,P)$ 4 time complexity and inefficient memory usage. In practical terms, the method is described as usable up to a few thousand points before time and memory become bottlenecks (Avinash et al., 21 Jul 2025, Siddique et al., 2018).

Fast-VAT addresses these limitations through two Python-level acceleration strategies. The first uses Numba’s Just-In-Time compilation, with core distance-matrix and reordering loops decorated by @numba.jit(nopython=True), compiling Python and NumPy loops to LLVM machine code. The second uses Cython with static typing, memoryviews, flattened arrays for cache locality, and malloc/free for index management. The reported result is up to $R=D(P,P)$ 5 speedup over the baseline implementation while preserving output fidelity; the paper states that all implementations produce identical reordered matrices and hence identical VAT images. On seven benchmark datasets, the reported Cython speedups included $R=D(P,P)$ 6 on Iris, $R=D(P,P)$ 7 on Mall Customers, and values around $R=D(P,P)$ 8– $R=D(P,P)$ 9 on several synthetic datasets (Avinash et al., 21 Jul 2025).

For larger-scale use, the same implementation study notes that storing a full $D^\ast_{ij}=D_{\pi[i],\pi[j]}$ 0 matrix may exceed available RAM for $D^\ast_{ij}=D_{\pi[i],\pi[j]}$ 1. The reported remedies are subsampling via sVAT, computing distances on the fly in blocks or on GPU, and offloading inner loops to CUDA via libraries such as CuPy or RAPIDS cuML. The same paper positions VAT as a quick pre-clustering diagnostic for interactive workflows when sub-second execution becomes feasible on $D^\ast_{ij}=D_{\pi[i],\pi[j]}$ 2– $D^\ast_{ij}=D_{\pi[i],\pi[j]}$ 3 (Avinash et al., 21 Jul 2025).

7. Limitations and methodological role

VAT is not presented in these sources as a universally reliable clustering algorithm in its own right. Its primary function is visual assessment of cluster tendency, even though it can be paired with MST cutting and single-linkage extraction. This distinction matters because the quality of the visualization depends directly on the dissimilarity representation and on the presence or absence of noise, bridges, and high-dimensional distortions (Rathore et al., 2020).

Several failure modes are explicit. VAT and iVAT are sensitive to noise and bridge points between clusters, and in such cases the corresponding images are often inconclusive. Bare VAT may fail on “tough” data where noise or chain-bridges spoil the diagonal block structure. Long, thin blocks or chains may indicate chain-shaped or non-convex structures that VAT does not clearly resolve. These limitations motivate iVAT’s path-based minimax transform and ConiVAT’s combination of metric learning, must-link/cannot-link constraints, and minimum-transitive-dissimilarity preprocessing (Rathore et al., 2020, Siddique et al., 2018).

A second limitation concerns representation in complex high-dimensional data. For raw image pixels, flattening leads to very large $D^\ast_{ij}=D_{\pi[i],\pi[j]}$ 4, Euclidean distances become less meaningful, spatial structure is lost, and VAT/iVAT heat maps on CIFAR-10 or MNIST can become blurred. DeepVAT addresses this by replacing raw inputs with self-supervised embeddings followed by t-SNE and MMRS, and its ablation study reports clear degradation when SimCLR or t-SNE is removed (Mazumder et al., 2023).

A methodological implication, explicitly drawn in the audio study, is that VAT can be used early in a pipeline as a model-agnostic exploratory tool to verify whether the chosen features encode sufficient separation for the intended task, to explain confusions observed in supervised classifiers, and to raise questions about unlabeled data before substantial effort is spent on downstream modeling. This suggests a role for VAT not only in cluster discovery but also in diagnosing when a desired label set is poorly aligned with the structure present in the data representation (Heise et al., 2021).