shankarpandala/lazypredict

What it solves

Lazy Predict simplifies the initial phase of machine learning model selection by allowing users to quickly benchmark a large number of models without writing extensive boilerplate code or performing manual parameter tuning. It eliminates the tedious process of trial-and-error when deciding which algorithm (classification, regression, or time series forecasting) is best suited for a specific dataset.

How it works

The library provides high-level wrappers (LazyClassifier, LazyRegressor, and LazyForecaster) that take training and testing data and automatically fit and evaluate a wide array of built-in models. It handles categorical encoding (OneHot, Ordinal, Target, Binary) and, for time series, automatically detects seasonal periods via ACF. Users can enable GPU acceleration for supported models (like XGBoost, CatBoost, and PyTorch-based LSTMs) to speed up the process.

Who it’s for

It is designed for data scientists and ML engineers who want to rapidly prototype and identify the most promising model architectures before committing to deep tuning and optimization.

Highlights

• Extensive Model Library: Over 40 built-in models across classification and regression, and 20+ forecasting models.
• Comprehensive Forecasting: Supports statistical models (ARIMA, ETS), ML models, deep learning (LSTM, GRU), and foundation models (TimesFM).
• Hardware Acceleration: GPU support via CUDA for boosting libraries, RAPIDS cuML, and deep learning models.
• Integrated Tooling: Built-in MLflow integration for experiment tracking and support for custom evaluation metrics.
• Flexible Preprocessing: Multiple categorical encoding strategies and configurable cross-validation.