uber/manifold

Provides interactive performance segmentation via k-means clustering to isolate underperforming data subsets, then visualizes feature distribution divergence between segments to surface root causes. Accepts classification and regression models through a standardized JSON format (features, predictions, ground truth) and renders dual comparison views: performance distributions across automatically-derived segments and ranked feature attribution heatmaps/histograms with KL-divergence scoring.

1,672 stars. No commits in the last 6 months.