LightGBM

Gradient-boosted decision trees remain a default for tabular problems, and LightGBM trains them quickly by bucketing continuous features into histograms, growing each tree leaf by leaf toward the split with the largest loss reduction, and partitioning categorical columns directly instead of one-hot encoding them. Objectives include Huber, Poisson, quantile and Tweedie regression, binary and multiclass classification, and LambdaRank for ranking, with metrics like AUC and NDCG, early stopping and L1 or L2 regularization. Training can spread across machines with feature, data or voting parallel modes, with an optional MPI build, and GPU builds exist for OpenCL, CUDA and ROCm. You get a C++ core and CLI, a Python package on pip and conda with scikit-learn style estimators and Dask support, an R package, a SWIG-built Java wrapper, and a C API that community bindings for .NET, Rust and Ruby build on. In March 2026 the repository moved from Microsoft to the lightgbm-org organization under the same maintainers. MIT licensed and free for commercial work, it suits data scientists with large tabular datasets.