🧠 What is MMM-Fair?

MMM-Fair is a fairness-aware machine learning framework designed to support high-stakes AI decision-making under competing fairness and accuracy demands. The three M’s stand for: • Multi-Objective: Optimizes across classification accuracy, balanced accuracy, and fairness (specifically, maximum group-level discrimination). • Multi-Attribute: Supports multiple protected groups (e.g., race, gender, age) simultaneously, analyzing group-specific disparities. • Multi-Definition: Evaluates and compares fairness under multiple definitions—Demographic Parity (DP), Equal Opportunity (EP), and Equalized Odds (EO).

MMM-Fair enables developers, researchers, and decision-makers to explore the full spectrum of possible trade-offs and select the model configuration that aligns with their social or organizational goals.

To Learn more about fairness-aware AI/ML please refer to the following materials:

1. Tutorial of Beginner's introduction to fair-ML: https://www.arjunroy.info/tutorials
2. Fair-ML Book: https://fairmlbook.org/pdf/fairmlbook.pdf

💬 No coding required:

MMM-Fair comes with an intuitive chat-based web UI (mmm-fair-chat) that guides users step by step—just like a human assistant would. You don’t need to write a single line of code. Simply upload your dataset (or use a built-in UCI dataset), select your fairness preferences, and explore trade-offs through automatically generated visual reports and summaries.

🧾 LLM-Powered Chart Explanations (New!)

Starting from v2.0.0, MMM-Fair supports automatic explanation of performance and fairness trade-off plots using LLMs (GPT, Groq, TogetherAI)

MMM-Fair is not just for developers, but also for policymakers, fairness auditors, and non-technical users.

Installation

pip install mmm-fair

Requires Python 3.11+.

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

Optional Installation for LLM enabled explanation

To enable this feature, install with extras (MMM-Fair currently supports only OpenAI (chatgpt), GroqAI, and TogetherAI but in future we plan to add more):

# OpenAI support
pip install "mmm-fair[llm-gpt]"

# Or for Groq
pip install "mmm-fair[llm-groq]"

# Or for Together.ai
pip install "mmm-fair[llm-together]"

#Or install all of them and later decide which one to use
pip install "mmm-fair[llm-gpt,llm-groq,llm-together]"

we do not provide any API keys for these models and to use the llm-explanation one needs to get their own api keys from the respective llm provider.

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

We provide two main classifiers:

1. MMM_Fair (Original Adaptive Boosting version)
2. MMM_Fair_GradientBoostedClassifier or MMM_Fair_GBT (Histogram-based Gradient Boosting approach) [recommended]

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.

Usage Overview

The mmm-fair package provides two different usage possibilities. One is a chat based on a web-based UI (specially tailored new user, with even non-technical abckground), and the other is command line based (for ML scientist, engineers, etc.)

Chat-Based

Right now the launch of the chat app is still terminal dependent (soon will release a destop app). So after installing the mmm-fair package one needs to bash in commandline:

mmm-fair-chat

and then in any browser copy paste:

http://127.0.0.1:5000

Then start chating with the interactive web app to get your MMM-Fair AI model.

(Optional) If you have installed MMM-Fair with LLM support and provide your API key during the session, the assistant can explain trade-off plots in natural language.

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,      # or max_iter=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.

Gradient-Boosted Style (recommended)

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)

📥 In-built Data Loader for UCI Datasets

MMM-Fair includes utility functions to seamlessly work with datasets from the UCI Machine Learning Repository.

🔧 Load a UCI dataset (e.g. Adult dataset)

from mmm_fair import data_uci
from mmm_fair import build_sensitives

# Load dataset with target column
data = data_uci(dataset_name="Adult", target="income")

🛡️ Define Sensitive Attributes

saIndex, saValue = build_sensitives(
    data.data,
    protected_cols=["race", "sex"],
    non_protected_vals=["White", "Male"]
)

🔀 Optional: Use with Train/Test Split

from sklearn.model_selection import train_test_split
import numpy as np

X = data.to_pred(sensitive=["race", "sex"])
y = data.labels["label"].to_numpy()
indices = np.arange(len(X))

X_train, X_test, y_train, y_test, id_train, id_test = train_test_split(
    X, y, indices, test_size=0.3, random_state=42, stratify=y
)

# Rebuild fairness attributes for the split sets
saIndex_train, saValue_train = build_sensitives(
    data.data.iloc[id_train], ["race", "sex"], ["White", "Male"]
)
saIndex_test, _ = build_sensitives(
    data.data.iloc[id_test], ["race", "sex"], ["White", "Male"]
)

✅ saIndex is a binary matrix (0 = protected, 1 = non-protected)

✅ saValue is a dictionary indicating protected status, e.g., {"race": 0, "sex": 0}

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.

Maintainer: Arjun Roy (arjunroyihrpa@gmail.com)

Contributors: Swati Swati (swati17293@gmail.com), Emmanoui Panagiotou (panagiotouemm@gmail.com)

🏛️ Funding

MMM-Fair is a research-driven project supported by several public funding initiatives. We gratefully acknowledge the generous support of:

Volkswagen Foundation – BIAS     EU Horizon – MAMMOth     EU Horizon – STELAR

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 arjun.roy@unibw.de Check out the source code at: GITHUB.