AdaBoost staged analysis with label-noise experiment

What this project does

This small teaching project demonstrates AdaBoost (ensemble boosting) using shallow decision-tree weak learners (stumps). It fits AdaBoostClassifier on a synthetic binary classification problem, inspects staged predictions across boosting rounds, and measures sensitivity to controlled label noise. The example emphasises how AdaBoost reweights examples and how that affects held-out error and the weights assigned to weak learners.

Data

The data are synthetic and created with scikit-learn’s make_classification. I build a moderately separable binary problem (800 rows, 10 numeric features). A controlled label-noise experiment is created by flipping a fixed fraction of training labels to simulate noisy annotation.

How to build it

1. Prepare the data — generate a synthetic binary dataset and make a stratified train/test split. Create a noisy version of the training labels by flipping a fixed percentage.
2. Build the experiment — train AdaBoostClassifier using DecisionTreeClassifier(max_depth=1) as weak learners. Train separate models on clean and noisy training labels, using the same random seed.
3. Evaluate the result — compute per-round held-out accuracy using staged predictions to inspect how error evolves with boosting rounds; extract estimator weights (alpha) from the fitted AdaBoost models.
4. Define a reviewable output — the run() function returns a JSON-serialisable dictionary containing final accuracies, best held-out accuracies and the boosting round where they were achieved, plus the weak-learner weights.

How to run

Run the module as a script: python project.py. The script prints a JSON object with the experiment metrics. The run(seed: int = 42) function can be imported and called from other code or tests.

How to test

This project includes pytest tests (test_project.py). Tests verify the expected keys, sensible numeric ranges, a conservative metric floor on held-out accuracy for the clean training case, and determinism (run(seed) produces the same output on repeated calls).

Results

Measured by running python project.py (seed 42) on 2026-10-23; the tests in test_project.py assert the floor.

Metric	Value
final_accuracy_clean	0.8550
best_accuracy_clean	0.8650
best_round_clean	17
final_accuracy_noisy	0.7950
best_accuracy_noisy	0.8100
best_round_noisy	43
noise_rate	0.2000
n_estimators	50

Also returned (see python project.py output): weak_weights (50 items).

Limitations

This uses synthetic data which is useful for teaching but not a substitute for a real domain dataset. The label-noise model is a simplistic simulation (random flips) and does not capture structured or adversarial noise. The experiment uses a single dataset split (with reproducible random seed); additional cross-validation would give a more robust assessment.

Next steps

• Extend to cross-validation or repeated splits to report mean and variance of metrics.
• Compare to bagging and random forests to highlight differences in how ensembles handle noisy labels.
• Visualise staged error curves and individual estimator decision boundaries for deeper intuition.

---

Full case study, build steps and evaluation checks: AdaBoost algorithm implementation on networkershome.com

Built by Networkers Home · Learn the skills in the AI Engineer Course · Part of 100 AI projects with code