AutoThink

Throw any data, get a working model.

One-click AutoML for tabular data.
Auto-detects task type • Engineers features • Trains LightGBM + XGBoost + CatBoost • Optimizes blend weights
All in a single function call.

---

Quickstart

pip install autothink

import pandas as pd
from autothink import fit

df = pd.read_csv("train.csv")
model = fit(df, target="price")
predictions = model.predict(pd.read_csv("test.csv"))

That's it. Three lines.

---

How It Works

Your DataFrame
     |
     v
+--------------------+     +-------------------------+     +---------------------+
| Task Detection     | --> | Intelligent             | --> | Adaptive Feature    |
| binary / multiclass|     | Preprocessing           |     | Engineering         |
| / regression       |     | missing values, encode, |     | learns thresholds & |
|                    |     | scale                   |     | interactions        |
+--------------------+     +-------------------------+     +---------------------+
                                                                    |
                                                                    v
+--------------------+     +-------------------------+     +---------------------+
| Verification       | <-- | Blend Optimization      | <-- | Ensemble Training   |
| fold stability,    |     | scipy-optimized weights  |     | LightGBM + XGBoost  |
| leakage check      |     | + Platt calibration      |     | + CatBoost (K-fold) |
+--------------------+     +-------------------------+     +---------------------+
     |
     v
  model.predict(test_df)

Step	What happens
Task detection	Determines binary, multiclass, or regression from the target column
Data validation	Checks for leakage, class imbalance, and quality issues
Preprocessing	Handles missing values, one-hot / target-encodes categoricals, scales numerics
Feature engineering	Learns optimal split thresholds and feature interactions from data
Ensemble training	Trains LightGBM, XGBoost, and CatBoost with adaptive hyperparameters
Blend optimization	Finds optimal ensemble weights via scipy on out-of-fold predictions
Calibration	Platt scaling for well-calibrated probabilities
Verification	Post-training diagnostics: fold variance, leakage, feature importance

---

Installation

From PyPI (coming soon):

pip install autothink

From source:

git clone https://github.com/ranausmanai/autothink.git
cd autothink
pip install -e .

With optional extras:

pip install autothink[dev]   # pytest
pip install autothink[api]   # FastAPI serving
pip install autothink[onnx]  # ONNX export

---

API Reference

fit(df, target, **kwargs)

One-line AutoML. Returns a fitted AutoThinkV4 instance.

Parameter	Type	Default	Description
df	DataFrame	required	Training data (features + target)
target	str	required	Name of the target column
time_budget	int	600	Maximum training time in seconds
verbose	bool	True	Log progress to console

AutoThinkV4

from autothink import AutoThinkV4

model = AutoThinkV4(time_budget=300, verbose=True)
model.fit(df, target_col="price")
preds = model.predict(test_df)

Attributes after fitting:

Attribute	Description
model.cv_score	Mean cross-validation score
model.cv_std	CV score standard deviation
model.task_info	Detected task type, metric, class info
model.verification_report	Post-training diagnostics

Logging

AutoThink uses Python's logging module. The library is silent by default.

import autothink
autothink.setup_logging()  # Enable INFO-level output to stderr

Or just use verbose=True (the default) which auto-configures a console handler.

---

Benchmarks

AutoThink V4 is competitive with FLAML and AutoGluon on standard tabular tasks:

Dataset	AutoThink V4	FLAML	AutoGluon
Heart Disease (AUC)	0.918	0.912	0.920
Loan Default (AUC)	0.874	0.869	0.871
House Price (RMSE)	30,241	31,102	29,876

_{60-second time budget, single 80/20 split, seed=42. Lower RMSE is better.}

---

Examples

See the examples/ directory:

Example	Description
quickstart.py	Minimal 15-line fit/predict on sklearn data
kaggle_competition.py	Full Kaggle pipeline with CLI and submission output
benchmark.py	Compare AutoThink against FLAML

---

Project Structure

autothink/
  __init__.py            # Public API: fit(), setup_logging()
  core/
    autothink_v4.py      # Main engine (TaskDetector, IntelligentEnsemble, AutoThinkV4)
    autothink_v3.py      # V3 engine (Kaggle-optimized)
    autothink_v2.py      # V2 engine (meta-learning)
    preprocessing.py     # IntelligentPreprocessor, FeatureEngineer
    feature_engineering_general.py  # Adaptive, data-driven feature engineering
    validation.py        # DataValidator, LeakageDetector
    meta_learning.py     # MetaLearningDB, dataset fingerprinting
    production.py        # ModelExporter, ModelCard, DriftDetector, APIGenerator
    advanced.py          # CausalAutoML, ExplanationEngine, SmartEnsemble
    kaggle_beast.py      # Competition-grade ensemble mode
    kaggle_fast.py       # Fast Kaggle mode
tests/                   # 25 tests (pytest)
examples/                # Quickstart, Kaggle, benchmark

---

Contributing

Contributions are welcome! Please open an issue or submit a PR.

# Development setup
git clone https://github.com/ranausmanai/autothink.git
cd autothink
pip install -e ".[dev]"
pytest tests/

---

License

Apache 2.0 — see LICENSE.

---

_{Built with scikit-learn, LightGBM, XGBoost, and CatBoost.}