Dataset shift characterization · Curated studies

dashi HORIZON

A curated repository of dataset shift characterizations across time and sources

Characterize covariate shift, monitor prior shift and track conditional changes using data distributions, non-parametric statistical manifolds and customizable machine learning models, built with the dashi Python package.

Browse studies
Across time 2010 2024
Across sources Source A Source B Source C Source D
Distribution-based

Unsupervised characterization

Builds temporal and source maps from data distributions and projects them onto non-parametric statistical manifolds to characterize dataset shifts.

Temporal & source maps IGT projections MSV metrics
train batch test batch
Model-based

Supervised characterization

Trains and evaluates different families of machine learning models on temporal or multi-source batches to characterize variability in performance.

Multi-batch evaluation Performance matrices Customizable models
Powered by the dashi Python package
pip install dashi