FeatureIQ

A knowledge graph and meta-learning framework for context-aware automated feature engineering.

The Problem

Feature engineering is the most impactful yet most time-consuming step in building machine learning models. Practitioners repeatedly make the same decisions -- log-transforming skewed features, one-hot encoding categoricals, scaling for distance-based models -- but these decisions are rarely captured in a reusable, systematic way. The right transformations depend on three interacting factors: the data characteristics, the problem type, and the downstream algorithm. Getting this wrong wastes time and hurts model performance.

FeatureIQ solves this by encoding feature engineering expertise into a structured ontology of rules, each backed by real citations from ML literature. It profiles your dataset, matches applicable rules based on column statistics and your declared algorithm, and builds an sklearn-compatible transformation pipeline automatically. For advanced users, a meta-learning layer trained on OpenML benchmarks can re-rank recommendations based on what historically worked for similar datasets.

Installation

pip install featureiq

With meta-learning support (LightGBM + OpenML):

pip install featureiq[meta]

Or with Poetry:

poetry add featureiq
poetry add featureiq -E meta  # with meta-learning

Quick Start

import pandas as pd
from sklearn.datasets import load_iris
from featureiq import FeatureIQ

data = load_iris()
X = pd.DataFrame(data.data, columns=data.feature_names)
y = pd.Series(data.target, name="target")

fiq = FeatureIQ(
    problem_type="multiclass_classification",
    algorithm="random_forest",
)
X_transformed = fiq.fit_transform(X, y)

# Inspect recommendations
print(fiq.score_report())
for col, recs in fiq.recommend().items():
    for rec in recs:
        print(f"  {col}: {rec['transformation']} (confidence: {rec['confidence']})")

FeatureIQ works inside sklearn pipelines:

from sklearn.pipeline import Pipeline
from sklearn.linear_model import LogisticRegression

pipe = Pipeline([
    ("features", FeatureIQ(problem_type="binary_classification", algorithm="logistic_regression")),
    ("clf", LogisticRegression()),
])
pipe.fit(X_train, y_train)
predictions = pipe.predict(X_test)

Key Features

Composability

Control exactly what FeatureIQ does:

fiq = FeatureIQ(
    problem_type="regression",
    algorithm="ridge",
    exclude_columns=["id", "timestamp"],           # pass through untouched
    override_transforms={"price": ["log_transform", "standard_scaler"]},  # your rules
    pre_transforms=[("custom_step", MyTransformer())],   # before FeatureIQ
    post_transforms=[("pca", PCA(n_components=10))],     # after FeatureIQ
)

Validation and Ablation

Verify that transforms actually help:

from sklearn.linear_model import Ridge

fiq.fit(X_train, y_train)
results = fiq.validate(X_train, y_train, estimator=Ridge(), cv=5, scoring="r2")
print(f"FeatureIQ: {results['featureiq_score']:.3f}  Baseline: {results['baseline_score']:.3f}")

# Per-transform impact analysis
ablation = fiq.ablate(X_train, y_train, estimator=Ridge(), cv=3)
for item in ablation[:5]:
    print(f"  {item['column']}/{item['transformation']}: impact={item['impact']:.4f}")

Cross-Column Awareness

FeatureIQ detects multicollinearity, computes VIF scores, and recommends accordingly for linear models -- while correctly skipping these checks for tree-based models that handle correlation natively.

Explainability

Every recommendation cites a real source:

fiq.explain_report()  # rich-formatted table with evidence and citations

How It Works

FeatureIQ is built on three pillars:

1. Data Profiling -- Automatically analyzes every column: distribution shape, missing rates, cardinality, outliers, temporal structure, and cross-column correlations.

2. Ontology Rules -- A curated knowledge graph of feature engineering rules, each specifying when a transformation should (or should not) be applied. Rules are written in YAML, cite real ML literature, and can be extended by anyone.

3. Meta-Learning (optional) -- A LightGBM model trained on OpenML benchmarks that predicts which transformations will improve performance for datasets similar to yours. Falls back gracefully to ontology-only mode when no trained model is available.

Supported Algorithms

Algorithm	Family
XGBoost	Tree-based
LightGBM	Tree-based
Random Forest	Tree-based
Decision Tree	Tree-based
Gradient Boosting	Tree-based
Logistic Regression	Linear Model
Linear Regression	Linear Model
Ridge	Linear Model
Lasso	Linear Model
SVM	Distance-based
KNN	Distance-based
MLP	Neural Network
AdaBoost	Ensemble
Bagging	Ensemble

Supported Problem Types

Problem Type	Value
Binary Classification	binary_classification
Multiclass Classification	multiclass_classification
Regression	regression
Time Series Forecasting	time_series_forecasting
Anomaly Detection	anomaly_detection

API Overview

Method	Description
fit(X, y)	Profile data, generate recommendations, build pipeline
transform(X)	Apply fitted pipeline to new data
recommend()	Return human-readable recommendations dict
validate(X, y, estimator)	Cross-validate FeatureIQ pipeline vs baseline
ablate(X, y, estimator)	Per-transform ablation analysis
explain(column=None)	Detailed explanations with evidence
explain_report()	Rich-formatted explanation table
score_report()	Plain-text summary of recommendations
get_pipeline()	Return the fitted sklearn Pipeline

Contributing

Contributions are welcome! See CONTRIBUTING.md for development setup and guidelines.

To contribute ontology rules, see docs/contributing_rules.md.

Citation

@software{featureiq2026,
  title={FeatureIQ: Context-Aware Automated Feature Engineering},
  author={Pratik Khedekar},
  year={2026},
  url={https://github.com/khedekarpratik0337/featureiq},
}

License

FeatureIQ is licensed under the Apache License 2.0.