lightgbm-org/LightGBM

Features

• Ensemble Methods - Implements a high-performance gradient boosting decision tree ensemble using leaf-wise growth and histogram binning.
• Gradient Boosting - Implements gradient boosting algorithms for high-performance classification, regression, and ranking tasks.
• Histogram-Based Learning - Discretizes continuous feature values into integer bins to reduce memory and accelerate split calculations.
• Distributed Training - Distributes the learning process across multiple machines to handle large-scale datasets.
• Distributed Training - Parallelizes model training using feature, data, or voting strategies across multiple nodes.
• Gradient Boosting Libraries - Provides a framework for constructing gradient-boosted decision tree ensembles that can be trained across multiple nodes.
• GPU-Accelerated Training - Offloads heavy feature binning and split search operations to CUDA or OpenCL devices.
• Large Scale Training - Processes massive datasets by distributing the learning workload across clusters to reduce training time.
• Leaf-wise Tree Growth - Expands trees by splitting the node with the largest loss reduction to achieve faster convergence and lower error.
• C-Based Engines - Implements the high-performance training and prediction engine in C for maximum execution speed.
• Native Categorical Splitting - Optimizes partitions for categorical variables using native splitting instead of one-hot encoding.
• Feature Contribution Analysis - Calculates SHAP values to determine the specific contribution of each feature to a final prediction.
• Feature Importance Attribution - Determines feature influence on predictions using SHAP values and feature contribution estimates.
• Gradient-Based Sampling - Implements gradient-based one-sided sampling to speed up training by prioritizing instances with large gradients.
• Learning to Rank Frameworks - Groups training records into query sets to enable learning-to-rank objectives.
• Model Evaluation Metrics - Calculates accuracy and error using specific evaluation metrics to monitor training progress and model quality.
• Distributed Machine Learning Integrators - Integrates with Dask to execute training and prediction tasks across distributed data collections.
• Model Deployment - Facilitates the loading of trained models into production environments for high-speed inference.
• Early Stopping Monitors - Tracks validation metrics and triggers early stopping to prevent overfitting during training.
• Distributed Execution - Spreads computations across multiple machines or GPUs to process massive datasets via parallel execution.
• Model Performance Optimization - Implements techniques like leaf-wise growth and regularization to improve prediction accuracy and prevent overfitting.
• Model Predictions - Generates predictions from trained models including raw scores, transformed scores, and leaf indices.
• Model Exporting - Exports trained models into C++ code or standalone files for production deployment without a runtime.
• Regularization and Sampling - Implements L1/L2 regularization, bagging, and column sub-sampling to prevent overfitting during the training process.
• Distributed Learning Communication - Synchronizes training state across multiple machines using socket or MPI communication.
• Language Wrappers - Provides high-level Python and R wrappers to simplify model configuration and data handling.
• MPI Communication - Uses MPI communication protocols to synchronize parallel training tasks across multiple machines.
• Model - Translates trained decision tree structures into C++ if-else statements for high-speed deployment.
• Regularized Tree Pruning - Applies L1 and L2 regularization during tree construction to prevent overfitting.
• Machine Learning - A fast, distributed, high performance gradient boosting framework.

Star history