py-evoFE v0.3.0 Released

To address the limitations of manual feature engineering in tabular data ML, the open-source library py-evoFE v0.3.0, which applies Genetic Programming, has been released. This tool automatically explores and optimizes complex ratios, nested grouped aggregations, and nonlinear dimension projections that GBDT models such as LightGBM or XGBoost struggle to discover on their own.

Key Features and Performance

py-evoFE explores the space of possible feature recipes using the following techniques:

• Hierarchical Chaining: Evolved features are used as components for the next generation, accumulating complex transformations.
• 40+ Built-in Transformers: Includes nonlinear arithmetic, target encoding (WoE, quantization, etc.), string similarity, manifold dimension reduction (PCA, UMAP, etc.), and graph and density clustering.
• High-Performance Computation: Accelerated by vectorized operations based on Polars and PyArrow. Additionally, matrix hashing and nearest-neighbor caching eliminate redundant calculations across CV folds.
• Multi-Fidelity Screening: Maximizes efficiency by rapidly screening the initial population at low fidelity and evaluating only promising candidates at full fidelity.

Parallel Search and Ensembling

By applying the Island Model, it supports multi-population parallel search across various topologies such as Ring, Torus, and Grid, exchanging information between populations via Gibbs migration. After the search concludes, the Greedy Caruana ensemble technique is applied to construct the final model.