LightGBM

Microsoft's gradient-boosting framework, built for training speed on very large datasets.

LightGBM trains gradient-boosted decision trees using a histogram-based approach: continuous values are bucketed into bins, which sharply cuts memory use and compute time. Leaf-wise tree growth, GOSS and EFB techniques for sampling gradients and bundling sparse features, and distributed learning make it a very fast, memory-efficient booster. It's our choice as soon as the data exceeds a million rows or hundreds of features and iterations need to stay short.