Nebius Science featured our recent work on QuadraSHAP - a fast and numerically stable algorithm for computing Shapley values.

Shapley values are a popular framework for explaining machine learning models, but their factorial computational cost makes them prohibitively slow in practice. Majid Mohammadi worked out the key theoretical insight - the painful sum over feature coalitions can be rewritten as a one-dimensional integral, solved efficiently by Gauss-Legendre quadrature. Grigory Reznikov turned this into fast, stable code for tree ensembles and kernels.

The result: 2-5x faster than the standard SHAP library on trees, and on kernels the gap widens further - 25x faster at 500 features, 95x at 1’000, and past 2’000 features the previous best simply times out while QuadraSHAP finishes in seconds.

The preprint is available on arXiv.