mmm-fair 0.4.3

A multi-objective multi-fairness boosting classifier

MMM-Fair is a multi-objective, fairness-aware boosting classifier originally inspired by the paper: "Multi-fairness Under Class-Imbalance"

https://link.springer.com/chapter/10.1007/978-3-031-18840-4_21

---

The original algorithm targeted Equalized Odds (a.k.a. Disparate Mistreatment). This MMM-Fair implementation generalizes to multiple fairness objectives:

• Demographic Parity (DP)

• Equal Opportunity (EP)

• Equalized Odds (EO)

We further improve the approach by:

1. Flexible Base Learners: Any scikit-learn estimator (e.g. DecisionTreeClassifier, LogisticRegression, ExtraTreeClassifier, etc.) can be used as the base learner.
2. Fairness-Weighted Alpha: The boosting weight (alpha) accounts for fairness metrics alongside classification error.
3. Dynamic Handling of Over-Boosted Samples: Reduces excessive emphasis on specific samples once fairness goals are partially met.
4. Gradient Boosted Tree version

Two Approaches: AdaBoost-Style vs. Gradient-Boosted Trees

We provide two main classifiers:

1. MMM_Fair (Original Adaptive Boosting version)
2. MMM_Fair_GradientBoostedClassifier (Histogram-based Gradient Boosting approach)

Both handle multi-objective, multi-attribute, and multi-type fairness constraints (DP, EP, EO) but differ in how they perform the boosting internally. You can choose via the command line argument --classifier MMM_Fair or --classifier MMM_Fair_GBT.

---

Installation

pip install mmm-fair

Requires Python 3.11+.

Dependencies: numpy, scikit-learn, tqdm, pymoo, pandas, ucimlrepo, skl2onnx, etc.

---

Usage Overview (AdaBoost-Style)

You can import and use MMM-Fair (original version):

from mmm_fair import MMM_Fair 
from sklearn.tree import DecisionTreeClassifier

Suppose you have X (features), y (labels)

mmm = MMM_Fair(
estimator=DecisionTreeClassifier(max_depth=5),
constraints="EO",        # or "DP", "EP"
n_estimators=1000,
saIndex=...,            # shape (n_samples, n_protected)
saValue=...,            # dictionary {'prot_att_column_name': prot value}
random_state=42,
# other parameters, e.g. gamma, saIndex, saValue...
)

mmm.fit(X, y)
preds = mmm.predict(X_test)

Fairness Constraints

• constraints="DP" → Demographic Parity

• constraints="EP" → Equal Opportunity

• constraints="EO" → Equalized Odds

In all cases, pass the relevant saIndex (sensitive attribute array) and saValue (dictionary of protected group mappings) to MMM_Fair if you want it to track fairness for different protected attributes.

---

Usage Overview (Gradient-Boosted Trees)

We also provide MMM_Fair_GradientBoostedClassifier. This uses a histogram-based gradient boosting approach (similar to HistGradientBoostingClassifier) but includes a custom fairness loss to train and then multi-objective post-processing step to select the best pareto-optimal ensemble round. Example:

from mmm_fair import MMM_Fair_GradientBoostedClassifier

clf = MMM_Fair_GradientBoostedClassifier(
    constraint="EO",        # or "DP", "EP"
    alpha=0.1,              # fairness weight
    saIndex=...,            # shape (n_samples, n_protected)
    saValue=...,            # dictionary or None
    max_iter=100,
    random_state=42,
    ## any other arguments that the HistGradientBoostingClassifier from sklearn can handle
)
clf.fit(X, y)
preds = clf.predict(X_test)

---

Train & Deploy Script

This package provides a train_and_deploy.py script. It:

1. Loads data (from a known UCI dataset or a local CSV).
2. Specifies fairness constraints, protected attributes, and base learner.
3. Selects either the original MMM_Fair or the new MMM_Fair_GradientBoostedClassifier via --classifier MMM_Fair or --classifier MMM_Fair_GBT.
4. Trains with your chosen hyperparameters.
5. Optionally deploys the model in ONNX or pickle format.

Key Arguments

•	--classifier: MMM_Fair (original boosting) or MMM_Fair_GBT (gradient-based).
•	--constraint: e.g., DP, EP, EO.
•	--n_learners: Number of estimators (for either version).
•	--pos_Class: Specify the positive class label if needed.
•	--early_stop: True or False, relevant for the GBT approach to enable scikit-learn’s early stopping.
•	--base_learner: E.g. tree, lr, logistic, etc. (for the original MMM_Fair).
•	--deploy: 'onnx' or 'pickle'.
•	--moo_vis True: Optionally visualize multi-objective (3D) plots (accuracy, class-imbalance, multi-fairness) after training, opening a local HTML page with interactive charts.

Example command:

1. Original AdaBoost MMM_Fair:

using UCI library

python -m mmm_fair.train_and_deploy \
  --dataset Adult \
  --prots race sex \
  --nprotgs White Male \
  --constraint EO \
  --base_learner Logistic \
  --deploy onnx \
  --moo_vis True

using local "csv" data

python -m mmm_fair.train_and_deploy \
  --dataset mydata.csv \
  --target label_col \
  --prots prot_1 prot_2 prot_3 \
  --nprotgs npg1 npg2 npg3 \
  --constraint EO \
  --base_learner tree \
  --deploy onnx

2. Gradient-Boosted MMM_Fair_GBT:

python -m mmm_fair.train_and_deploy \
  --classifier MMM_Fair_GBT \
  --dataset mydata.csv \
  --target label_col \
  --prots prot_1 prot_2 \
  --nprotgs npg1 npg2 \
  --constraint DP \
  --alpha 0.5 \
  --early_stop True \
  --n_learners 100 \
  --deploy pickle \
  --moo_vis True

Note:

1. Setting --moo_vis True triggers an interactive local HTML page for exploring the multi-objective trade-offs in 3D plots (accuracy vs. class-imbalance vs. fairness, etc.).
2. Currently the fairness intervention only implemented for categorical groups. So if protected attribute is numerical e.g. "age" then for non-protected value i.e. --nprotgs provide a range like 30_60 as argument.

---

Additional options

If you want to select the best theta from only the Pareto optimal ensembles set (default is False and selects applies the post-processing to all set of solutions):

--pareto True

If you want to provide test data:

--test 'your_test_file.csv'

Or just test split:

--test 0.3

If you want change style (default is table, choose from {table, console}) of report displayed (Check FairBench Library for more details):

--report_type Console

When deploying with 'onnx', we change the models to ONNX file(s), and store additional parameters in a model_params.npy. This gets zipped into a .zip archive for distribution/analysis.

---

MAMMOth Toolkit Integration

For the bias exploration using MAMMOth pipeline it is really important to select 'onnx' as the '--deploy' argument. The ONNX model accelerator and model_params.npy are used to integrate with the MAMMOth-toolkit or the demonstrator app from the mammoth-commons project.

By providing the .zip archive, you can:

•	Upload it to MAMMOth,

•	Examine bias and performance metrics across subgroups,

•	Compare fairness trade-offs with a user-friendly interface.

---

Example Workflow

1. Choose Fairness Constraint: e.g., DP, EO, or EP.
2. Define sensitive attributes in saIndex and the protected-group condition in saValue.
3. Pick base learner (e.g., DecisionTreeClassifier(max_depth=5)) or gradient-based approach.
4. Train with a large number of estimators (n_estimators=300 or max_iter=300).
5. Optionally do partial ensemble selection with update_theta(criteria="all") or update_theta(criteria="fairness") .
6. Export to ONNX or pickle for downstream usage.
7. Use --moo_vis True to open local multi-objective 3D plots for deeper analysis.
8. Upload the .zip file (if exported to onnx) to MAMMOth for bias exploration.

---

References

“Multi-Fairness Under Class-Imbalance,” Roy, Arjun, Vasileios Iosifidis, and Eirini Ntoutsi. International Conference on Discovery Science. Cham: Springer Nature Switzerland, 2022.

License & Contributing

This project is released under [Apache License Version 2.0]. Contributions are welcome—please open an issue or pull request on GitHub.

Contact

For questions or collaborations, please contact [e-mail] Check out the source code at: GITHUB.