Can an algorithm recommend romantic matches fairly? A scoring system for a fictional dating app, evaluated through the lens of performance, interpretability, stability, and fairness.
HEC Match is a group project for the MSc DSAIB course "Interpretability, Stability, and Algorithmic Fairness" (Prof. Christophe Pérignon & Dr. Sébastien Saurin, HEC Paris, September 2026).
We built a recommendation engine for a fictional dating app using the Speed Dating Experiment dataset (Fisman & Iyengar, Columbia Business School, 2002–2004). The dataset contains 8,378 observations from 21 speed-dating sessions where participants rated each other across multiple dimensions.
Target variable: dec — whether participant A wants to see participant B again (~40% positive rate).
Core constraint: Only pre-date variables are used as features (questionnaire responses, demographics, declared preferences). Post-date ratings are excluded to avoid information leakage — resulting in a modest but honest AUC ceiling of ~0.60.
| Dimension | Insight |
|---|---|
| Performance | All 3 models converge to the same AUC plateau (~0.59–0.60 GroupKFold), confirming the problem's intrinsic difficulty — not a modeling failure |
| Interpretability | XGBoost top features: int_corr (shared interest correlation), attr3_1 (self-assessed attractiveness), fun1_1 (partner's sense of humor priority) |
| Stability | 16.7% (Logit) and 21.4% (XGBoost) of decisions flip across bootstrap retrainings at threshold 0.50 |
| Fairness | Models reproduce human racial preferences from the data — Asian candidates are under-exposed by +8.6 pp (societal bias). XGBoost adds ≤ 4 pp of algorithmic bias; Logit amplifies Latino discrimination by −12.5 pp |
| Mitigation | Removing 4 FPDP-identified racial proxies reduces TabICL's algorithmic bias on the "Other" group from +9.1 pp to +3.3 pp with negligible AUC loss |
match_HEC/
├── app/ # Interactive web app (static HTML, no server)
│ ├── index.html # Self-contained app — open in browser
│ ├── score.js # Client-side logit/XGBoost scoring
│ └── build.py # Regeneration script
│
├── colab/ # Google Colab scripts (GPU required)
│ ├── train_tabicl.py # TabICL baseline training
│ ├── train_tabicl_mitigated.py # TabICL mitigated training (154 features)
│ ├── predict_tabicl.py # Batch inference
│ └── xper_tabicl.py # XPER decomposition on Colab
│
├── data/
│ ├── clean.parquet # Engineered dataset (generated by build_dataset)
│ ├── features.json # Feature contract (158 XGB / 156 Logit)
│ ├── split.json # Train/test wave split
│ ├── tabicl_predictions.parquet # Pre-computed TabICL probabilities
│ ├── tabicl_mitigated_predictions.parquet
│ ├── mitigated_features.json # 154-feature contract (4 proxies removed)
│ └── economic_assumptions.json # P&L cost matrix (X=2€, Y=0.5€, Z=0€)
│
├── docs/
│ ├── soutenance_notes.md # Detailed defense notes (FR)
│ └── data_dictionary.md # Variable documentation
│
├── models/
│ ├── xgb.joblib # XGBoost baseline
│ ├── logit.joblib # Logistic regression baseline
│ ├── xgb_mitigated.joblib # XGBoost without racial proxies ⭐
│ ├── logit_mitigated.joblib # Logit without racial proxies
│ ├── tabicl.joblib # TabICL (CUDA-only, not portable)
│ └── tabicl_mitigated.joblib # TabICL mitigated (CUDA-only)
│
├── notebooks/
│ ├── 00_eda_cleaning.ipynb # EDA & data cleaning
│ ├── 02_interpretability.ipynb # SHAP, LIME, PDP/ICE, XPER
│ ├── 05_performance.ipynb # Model comparison & calibration
│ ├── 06_economics.ipynb # P&L analysis & threshold optimization
│ └── 07_tradeoffs.ipynb # Multi-criteria trade-off synthesis
│
├── reports/
│ ├── economics/ # P&L curves, sensitivity analysis
│ ├── fairness/ # TOST intervals, bias decomposition, mitigation
│ ├── interpretability/ # SHAP beeswarm, odds ratios, surrogate trees
│ ├── performance/ # Calibration, threshold grids, XPER
│ ├── stability/ # Bootstrap distributions, decision flips
│ └── tradeoffs/ # Final arbitrage tables
│
├── src/
│ ├── build_dataset.py # Feature engineering pipeline
│ ├── metrics.py # TOST, shared fairness metrics
│ ├── mitigation.py # FPDP proxy identification & removal
│ ├── generate_fairness_reports.py
│ ├── fairness_at_optimal_threshold.py
│ ├── fairness_bias_decomposition.py
│ ├── interpretability.py # SHAP, LIME, PDP, XPER, surrogate
│ ├── performance.py # All-model evaluation & calibration
│ ├── economics.py # P&L simulation & top-K strategies
│ ├── stability.py # Bootstrap retraining & decision flip analysis
│ ├── tradeoffs.py # Multi-criteria comparison table
│ ├── tabicl_model.py # TabICL wrapper
│ └── app_export.py # Export data for the web app
│
├── tests/
│ └── test_stability.py
│
├── 01_data_models_v0.py # End-to-end baseline script
├── 04_stability.ipynb # Stability notebook (root level)
├── requirements.txt # Core dependencies
├── requirements-stability.txt # Stability-specific dependencies
└── requirements-interpretability.txt
- Python 3.11+
- The raw dataset from Kaggle (Speed Dating Experiment)
# Clone the repository
git clone https://github.com/MaximilienLuc/match_HEC.git
cd match_HEC
# Create and activate virtual environment
python3 -m venv .venv
source .venv/bin/activate # macOS/Linux
# Install dependencies
pip install -r requirements.txtDownload Speed Dating Data.csv and Speed Dating Data Key.doc from Kaggle and place them in data/.
# 1. Build the clean dataset
python -m src.build_dataset
# 2. Train baseline models (Logit + XGBoost)
python 01_data_models_v0.py
# 3. Run fairness audit (TOST + bias decomposition)
python -m src.generate_fairness_reports
# 4. Run mitigation pipeline (FPDP proxy removal + retraining)
python -m src.mitigation
# 5. Run performance evaluation (all models)
python -m src.performance
# 6. Run economics analysis (P&L, threshold optimization)
python -m src.economics
# 7. Run stability analysis
python -m src.stability
# 8. Generate trade-off synthesis
python -m src.tradeoffs
# 9. Build the web app
python -m app.buildTabICL cannot run on CPU due to a known segfault in tabicl==2.2.0. To train or run inference:
- Open Google Colab with a GPU runtime (Edit > Notebook settings > GPU)
- Upload
data/clean.parquet,data/features.json, anddata/split.json - Run
colab/train_tabicl.py(baseline) orcolab/train_tabicl_mitigated.py(mitigated)
Pre-computed predictions are versioned in data/tabicl_predictions.parquet and data/tabicl_mitigated_predictions.parquet.
| Model | Type | AUC (test) | AUC (GroupKFold 5×) | Features |
|---|---|---|---|---|
| Logistic Regression | White-box | 0.589 | 0.603 ± 0.024 | 156 |
| XGBoost | ML (boosted trees) | 0.611 | 0.599 ± 0.024 | 158 |
| TabICL | Tabular Foundation Model | 0.632 | 0.591 ± 0.038 | 158 |
Note: The GroupKFold AUC (splitting by wave) is the reliable metric to cite. The single-split test AUC, particularly for TabICL, is likely optimistic due to high fold variance.
All three models have mitigated versions trained on 154 features after removing 4 FPDP-identified racial proxies (attr3_1_B, career_c_B_12, go_out_B, intel3_1_B).
Our fairness framework uses the TOST equivalence test (Schuirmann, 1987), which reverses the burden of proof: models are considered biased by default and must statistically prove that exposure gaps remain within ±10 percentage points.
Key result: All three models reproduce societal biases present in the raw data (Asian candidates are under-favored by +8.6 pp in humans' actual choices). XGBoost is the most algorithmically neutral (≤ 4 pp added bias on all groups), while Logit significantly amplifies discrimination against Latino candidates (−12.5 pp of pure algorithmic bias).
Open app/index.html in any browser — no server required. Features:
- Recommendations: For any test user, see their candidates ranked by the selected model with the top-3 picks highlighted and SHAP/coefficient explanations
- Simulator: Logit and XGBoost mitigated are scored client-side in JavaScript (
score.js), matching Python output to < 1e-4 - Model comparison: Interactive tables from
reports/tradeoffs/andreports/economics/
Regenerate after model changes: python -m app.build
- Fisman, R., & Iyengar, S. S. (20 06). Gender Differences in Mate Selection: Evidence from a Speed Dating Experiment. The Quarterly Journal of Economics.
- Hitsch, G. J., Hortaçsu, A., & Ariely, D. (2010). Matching and Sorting in Online Dating. American Economic Review.
- Schuirmann, D. J. (1987). A Comparison of the Two One-Sided Tests Procedure and the Power Approach for Assessing the Equivalence of Average Bioavailability. Journal of Pharmacokinetics and Biopharmaceutics.
- Pérignon, C., & Saurin, S. (2026). Interpretability, Stability, and Algorithmic Fairness. Course materials, HEC Paris MSc DSAIB.
This project is an academic exercise. The Speed Dating Experiment dataset is publicly available on Kaggle and OpenML under their respective terms.