See the pipeline. Run the model. Export the answer.
ML Clustering is a visual scikit-learn workbench for CSV files. Pick an algorithm, or describe the pipeline in English and let the copilot draw the graph. Run it in the browser, then keep the chart, the hold-out metrics, and the Python.
Sample tables such as Iris Flowers run with no account. Uploading your own file, saving the graph, and downloading exports starts on the free plan: 50 jobs / month, 100k rows, no credit card. Training scores are marked separately from hold-out scores, so the number you ship is the one measured on rows the model did not see.
What you can run
Each family has its own workbench. Preprocessing imputes, scales, and encodes before a model sees the table. The pipeline canvas chains those steps into one graph, and the example gallery shows the same algorithms on Iris, wine, and a synthetic customer demo, with the metrics written out.
Clustering
Group rows that belong together when you have no label column. Partition methods such as K-Means, density methods such as DBSCAN and HDBSCAN, and probabilistic models such as Gaussian mixtures each assume a different shape of cluster.
- K-Means Clustering
- K-Medoids Clustering
- DBSCAN Clustering
- HDBSCAN Clustering
- OPTICS Clustering
- Mean Shift Clustering
- Agglomerative Hierarchical Clustering
- BIRCH Clustering
- Bisecting K-Means Clustering
- Spectral Clustering
- Gaussian Mixture Model (GMM)
- Fuzzy C-Means Clustering
- Affinity Propagation Clustering
- Growing Neural Gas (GNG)
Classification
Predict a category from the other columns. Logistic regression, support vector machines, tree ensembles, and a small neural net all train on a labeled CSV and report hold-out accuracy, precision, recall, and a confusion matrix.
Regression
Predict a number. Linear models with L1 and L2 penalties sit next to tree ensembles and support vector regression, with R², MAE, and residual plots on the rows the model did not train on.
Dimensionality reduction
Compress many columns into a plot you can read. PCA and factor analysis keep linear structure. t-SNE and UMAP keep local neighborhoods, which is what you want when the clusters are curved.
Worked examples
These pages describe a finished run: the dataset, how the algorithm behaves on it, and which metric to read. Open one in the sandbox to rerun it on the same sample.
K-Means on Iris Flowers
Three species, four measurements, and a silhouette score you can recompute in the sandbox. Source: UCI Machine Learning Repository, Iris (Fisher, 1936). License: CC BY 4.0.
DBSCAN on Wine (UCI wine recognition)
Density clustering that chooses the number of groups and marks sparse rows as noise. Source: UCI Machine Learning Repository, wine recognition. License: CC BY 4.0.
K-Means on Mall Customers
A 200-row tutorial table, not a scale proof. Five clusters, seed 42. Source: Mall customer segmentation tutorial. License: Tutorial table hosted on this site.
HDBSCAN on Iris Flowers
The hierarchical version of DBSCAN, for clusters that are not all the same density. Source: UCI Machine Learning Repository, Iris (Fisher, 1936). License: CC BY 4.0.
Blueprints
A blueprint is a whole pipeline, already wired: loader, scaler, model, and plot. Load it on the canvas and press Run. The write-ups live under Pipeline blueprints.
- Customer Segmentation Pipeline — K-Means on the 200-row Mall Customers tutorial table. The loader, a standard scaler, k of 5, a plot, and the cluster metrics. A small segmentation example, not a scale proof.
- Telco Customer Churn Pipeline — Logistic regression on the 7,043-row IBM Telco Customer Churn sample. One-hot encoding, customer id left out, a stratified 20 percent hold-out, and the scores for customers who left.
- Telco Customer Segmentation Pipeline — K-Means on the 7,043-row IBM Telco sample. Customer id, gender, total charges, and churn stay out. Tenure and monthly charges are standardised. k of 2 is the silhouette peak from 2 through 8.
- Online Retail Customer Segmentation Pipeline — K-Means on 4,338 customers rolled up from the UCI Online Retail invoices. Log transform, standard scale, and k of 2, the silhouette peak from 2 through 8.
- Credit Card Fraud Pipeline — Logistic regression on a 100,000-row sample of the credit-card fraud table. Every fraud row is kept, Time stays out, the hold-out is stratified, and class weight is balanced.
- Density-Based Clustering Pipeline — DBSCAN on the synthetic customer-segments demo. Imputes missing values, scales features, and finds clusters without a preset count. The sample is generated for this site, not a production table.
How a run works
- Open a sample, or upload a CSV once you have an account. The Iris pipeline is the shortest path: scaler, K-Means with three clusters, and a plot.
- Tune the parameters in the sidebar, or ask the copilot to add a step. The graph stays visible, so you can see which node feeds the next one.
- Read the hold-out metrics, export labels or probabilities, and download the scikit-learn script that repeats the same steps outside the browser.
Plans and row limits are on the pricing page.
