Cluster Tabular Data
Paste a CSV, pick KMeans, DBSCAN, or hierarchical clustering, and get a colored scatter plot, a labeled CSV, or a JSON report with centroids and a silhouette score. Pure-Rust clustering runs in your browser — nothing is uploaded.
Cluster CSV data three ways, right in your browser
Paste a table, choose an algorithm, and see the groups your data falls into. Data Clusterer runs KMeans, DBSCAN, and hierarchical (agglomerative) clustering entirely in your browser — the CSV never leaves your machine, and there is no model to download. The result is a self-contained SVG scatter plot coloured by cluster (with centroid markers and a legend), a labeled CSV mapping every row to its cluster, or a JSON report with cluster sizes, centroids, and a silhouette quality score.
- KMeans — partitions the rows into a fixed number of clusters (
k) around their centroids. Deterministic farthest-point seeding means the same data always gives the same clusters. - DBSCAN — finds clusters by density using
eps(neighbourhood radius) andmin_samples. Points in no dense region are labeled noise (drawn in grey) — great for spotting outliers. - Hierarchical — repeatedly merges the closest clusters until the target count remains, with average, complete, single, or Ward linkage.
Feature columns are picked by header name or 1-based index — or auto-detected (every fully-numeric column). Turn on standardize (z-score) so columns on different scales (say, age vs. income) contribute fairly. Distance is Euclidean.
Worked example
Input (KMeans, k = 3, standardized):
height,weight
150,52
158,55
162,58
170,70
174,73
178,76
185,92
189,96
193,101
KMeans groups the nine rows into three body-size segments. Switch Output to JSON to read them back as data:
{
"method": "kmeans",
"points": 9,
"clusters": 3,
"noise": 0,
"normalized": true,
"features": ["height", "weight"],
"silhouette": 0.78,
"cluster_detail": [
{ "cluster": 1, "size": 3, "centroid": [156.667, 55] },
{ "cluster": 2, "size": 3, "centroid": [174, 73] },
{ "cluster": 3, "size": 3, "centroid": [189, 96.333] }
]
}
The silhouette score runs from −1 to 1 — higher means tighter, better-separated
clusters, so you can compare different k values or methods.
Reading the scatter plot
Each point is a row, coloured by its cluster; the larger diamond in each colour is that cluster's centroid. With exactly two feature columns the axes are those columns. With more than two features, the plot shows a 2-D PCA projection (PC1 / PC2) so the structure is still visible — the clustering itself always uses all selected columns. A single feature is plotted against its row number.
Limits
- Clustering runs on up to 10,000 rows for KMeans/DBSCAN and 1,500 rows for hierarchical (which needs an n×n distance matrix); up to 50 feature columns.
- Distance is Euclidean; there is no Manhattan/cosine option (KMeans centroids assume Euclidean geometry).
- The silhouette score is reported for up to 4,000 clustered points.
FAQ
Which algorithm should I choose?
Use KMeans when you already know roughly how many groups you want and they're
blob-shaped. Use DBSCAN when you don't know the count, the clusters may be irregular,
or you want outliers flagged as noise — tune eps and min_samples. Use hierarchical
for small datasets when you want control over how clusters merge (the linkage), or to
try several k values from one structure.
Should I turn on standardize (z-score)?
Usually yes. Standardizing rescales every feature to zero mean and unit variance, so a
large-magnitude column (e.g. income in the tens of thousands) doesn't dominate a
small-magnitude one (e.g. age). Turn it off when your columns are already on the same
scale and you want eps (for DBSCAN) to be interpreted in the raw units.
How do I pick DBSCAN's eps and min_samples?
eps is the maximum distance for two points to be neighbours; min_samples is how many
neighbours (including the point itself) a point needs to anchor a cluster. If everything
collapses into one cluster, lower eps; if everything is labeled noise, raise eps or
lower min_samples. With standardize on, eps is in standard-deviation units, so values
around 0.3–1.0 are a good starting range.
What is the silhouette score?
It measures how well each point sits in its cluster versus the nearest other cluster, averaged over all points, from −1 to 1. Values near 1 mean tight, well-separated clusters; near 0 means overlapping clusters; negative means points may be in the wrong cluster. It's shown in the JSON output so you can compare settings objectively.
My data has more than two columns — how is it plotted?
Clustering always uses every selected feature column. For the scatter plot, two features are drawn directly on the axes; with three or more features the tool computes a 2-D principal-component projection (PC1 vs PC2) so you can still see the cluster structure. Use the CSV or JSON output to get the exact per-row labels and centroids.
Is my data uploaded anywhere?
No. All parsing, clustering, PCA, and rendering happen locally in your browser via WebAssembly. Nothing is sent to a server and there is no model download.
Developer & Automation Access
Run it from the terminal
Same engine as this page, headless — via the gizza CLI:
gizza tool data-clusterer "height,weight
150,52
160,61
165,64
170,70
182,84
188,92
191,99"New to the CLI? Get gizza →
Open it by URL
Pre-fill and auto-run this tool with query parameters — the names match the API/CLI:
https://gizza.ai/tools/data-clusterer/?data=height%2Cweight%0A150%2C52%0A160%2C61%0A165%2C64%0A170%2C70%0A182%2C84%0A188%2C92%0A191%2C99&method=kmeans&clusters=3&eps=1.0&min_samples=4&linkage=average&columns=height%2Cweight&normalize=true&output=chart&title=Customer%20segments&width=700&height=500Machine-readable descriptor: tool.json — title + parameters JSON Schema for agents.