OpenFE
Gradient-boosted trees are very good at using features and very bad at inventing them. If the signal lives in price per kilogram, or in a customer's order count per month, somebody has to write that column. OpenFE automates that step for tabular data, and does it well enough to compete with experienced human feature engineers.
Generate wide, then prune hard
OpenFE applies arithmetic and aggregation operators across the raw columns — ratios, differences, products, group-by statistics, frequency counts, combinations of categorical values — which turns a modest table into tens of thousands of candidate features. The interesting part is the filtering, because retraining a model per candidate would be hopeless. OpenFE scores a candidate with FeatureBoost: it holds the base model's predictions fixed as a starting point and measures only the extra improvement the new feature buys, so one cheap boosting step yields an incremental value. A successive-halving pass over growing data subsets then discards weak candidates on little data and spends real compute only on survivors.
Using it
The interface is deliberately plain: fit on the training frame and the label, get back a ranked list of generated features, then transform train and test together and hand the result to LightGBM or XGBoost. Feeding the top dozen is usually right — wholesale injection adds noise and raises the risk of leakage.
Where it sits
Unlike Featuretools, which synthesises features by walking relationships between several tables, and tsfresh, which extracts statistics from time series, OpenFE targets a single flat table and the operator space a practitioner would try by hand.
OpenFE is an automated feature-generation tool for tabular data. It applies arithmetic and aggregation operators across raw columns to produce tens of thousands of candidate features, then prunes them cheaply: FeatureBoost measures only the incremental gain a candidate adds on top of the base model's predictions, and a successive-halving pass discards weak candidates on small data subsets. The surviving features feed a gradient-boosting model and rival hand-crafted feature engineering.
A good cook knows that two ingredients already in the kitchen can combine into something far better than either alone. Prediction models are similar: the raw columns in a spreadsheet are often less useful than the ratios, differences and counts you can build out of them. OpenFE invents those combinations automatically. It mixes the existing columns in thousands of ways, then runs a fast elimination contest to find the handful that genuinely improve predictions, throwing out the rest. The result is the kind of extra column an experienced data scientist would have added by hand — found in minutes instead of days.