Plotly-native clustering diagnostics.
Functions:
elbow_curve — sum-of-squared-distances vs. k for KMeans-family
models.silhouette_curve — silhouette score vs. k.silhouette_plot — per-sample silhouette for the chosen model.cluster_distribution — bar of cluster sizes.embedding_2d — t-SNE / UMAP / PCA scatter colored by cluster.The screen-side composition: the "Clustering" tab in the Run-detail view shows elbow + silhouette side-by-side for tuning k, then embedding_2d for the chosen model.
Sum-of-squared-distances vs. k. The "elbow" is where adding clusters stops meaningfully reducing inertia — pick k there.
Works on any clusterer that accepts n_clusters and exposes
inertia_ after fit (KMeans-family).
Mean silhouette score vs. k. Higher is better; ranges -1..1.
Per-sample silhouette grouped by cluster label.
A good clustering shows similar widths across clusters and high average silhouette. Negative silhouettes indicate likely misclassification.
Bar chart of how many samples fall into each cluster.
2-D embedding of X colored by the clusterer's labels.
Methods:
pca — sklearn PCA (fast, deterministic, no extra dep).tsne — sklearn TSNE (slower, captures non-linear structure).umap — only if umap-learn is installed.