Tag: #gradient-boosting5 results
Meta open-sources MCGrad, a Python package for multicalibration using gradient boosted decision trees. It corrects miscalibration in subgroups while scaling to large datasets. Presented at KDD 2026 with strong results on 100+ production models.
PerpetualBooster v1.9.0 is a Rust-based GBM using a single budget parameter instead of hyperparameter tuning, now featuring causal ML, drift detection, and conformal prediction. It supports 19 objectives, multi-output, and continual learning. Benchmarks show it matching LightGBM accuracy with 405x speedup and beating AutoGluon.
A data science team shares findings on how a top-performing LightGBM feature based on Bayesian target encoding led to worse predictions due to overfitting on irreducible label variance. The article highlights the dangers of target encoding leakage in gradient boosting models.
ibu-boost is a new gradient-boosted tree library applying a screening transform to absolutely reject poor splits, eliminating the need for min_gain_to_split tuning. It supports non-oblivious and oblivious trees, MSE/binary loss, missing values, and Triton GPU kernels for speedups. Benchmarks on California Housing show 12% RMSE gap to LightGBM but 3x GPU acceleration.
A new study evaluates whether replacing continuous predictors with guideline-based clinical categories can make ischaemic stroke outcome models more interpretable. Across three treatment cohorts, categorised gradient-boosted models performed comparably in two cohorts, while accuracy declined significantly in one.