FA-BUHS: Fairness-Aware Biased User History Synthesis

This repository contains the official PyTorch implementation of FA-BUHS (Fairness-Aware Biased User History Synthesis), an advanced long-tail recommendation framework.

FA-BUHS introduces a paradigm shift from post-hoc correction to upstream fairness intervention. By actively synthesizing a fairness-constrained interaction history, it mitigates popularity and demographic biases at their source before latent embeddings are formed. It further utilizes conditional adversarial debiasing to scrub demographic leakage from representations while strictly preserving collaborative filtering signals.

✨ Key Features

• Upstream Fairness Intervention: Modifies the user interaction history prior to representation learning, acting as a structural causal intervention.
• Fairness-Constrained Sampling: Generates synthetic interaction histories that balance long-tail exposure with demographic group representation.
• History Synthesis Variants: Supports multiple aggregation mechanisms including Mean, Attention, and GRU-based temporal modeling.
• Conditional Adversarial Debiasing: Employs Gradient Reversal Layers (GRL) conditioned on the interaction label to remove sensitive attributes without destroying utility.
• Post-Hoc Fairness Re-ranking: (Optional) A lightweight exposure-aware re-ranking step for inference-time fairness guarantees.
• Built-in Evaluation: Computes Position-Discounted Gini index, Demographic Parity, and Exposure Fairness alongside standard recommendation metrics (HR@K, NDCG@K).

---

📂 Repository Structure

├── main.py                 # Main training and evaluation loop
├── fairness_utils.py       # Core fairness logic (Sampling, Metrics, GRL, Re-ranking)
├── model/
│   ├── __init__.py         # Model configurations and registry
│   └── sparsenn.py         # Two-tower Neural architectures and Synthesis modules
├── data/
│   ├── dataloader.py       # PyTorch Dataset and Collator classes
│   ├── datareader.py       # Raw dataset parsers (ML-1M, BookCrossing)
│   └── structs.py          # Interaction graph structures and data splits
└── README.md

---

⚙️ Installation & Requirements

Clone the repository:

git clone [https://github.com/prateek4ai/fairnessawarebuhs.git](https://github.com/prateek4ai/fairnessawarebuhs.git)
cd fairnessawarebuhs

Install the required dependencies:

pip install torch numpy pandas scipy sentence-transformers wandb absl-py tqdm

---

🚀 Usage

The framework is highly configurable via command-line flags powered by absl-py. Experiment tracking is integrated natively with Weights & Biases (wandb).

Running the Baseline (Standard BUHS)

To run the standard long-tail synthesis without fairness constraints on MovieLens-1M:

python main.py \
    --dataset_dir="./datasets" \
    --dataset="ml" \
    --enable_fairness=False \
    --synthesis_type="gru" \
    --sample_size=15 \
    --epochs=50 \
    --batch_size=2048 \
    --device="cuda"

Running FA-BUHS (Fairness-Enabled)

To enable the full FA-BUHS pipeline (Fairness-Constrained Sampling + Conditional Adversarial Debiasing):

python main.py \
    --dataset_dir="./datasets" \
    --dataset="ml" \
    --enable_fairness=True \
    --fairness_sampling=True \
    --synthesis_type="gru" \
    --sample_size=15 \
    --adversarial_lambda_user=0.1 \
    --adversarial_lambda_item=0.1 \
    --fairness_weight=0.5 \
    --epochs=50 \
    --device="cuda"

Applying Inference-Time Fairness Re-ranking

You can apply the optional fairness-aware re-ranking during the evaluation phase:

python main.py \
    --dataset="ml" \
    --enable_fairness=True \
    --rerank_fair=True \
    --rerank_alpha=0.2 \
    --exposure_floor=0.1

Flag	Type	Default	Description
--dataset	str	"ml"	Dataset to use (ml for MovieLens, bx for BookCrossing).
--enable_fairness	bool	False	Enable the overarching fairness framework.
--fairness_sampling	bool	False	Enable upstream fairness-constrained sampling.
--synthesis_type	str	"mean"	User history synthesis method (mean, attention, or gru).
--sample_size	int	15	Synthesis horizon (N). Optimal recommended value is 15.
--adversarial_lambda_user	float	0.1	Weight for the user-side conditional adversarial loss (λ_u).
--adversarial_lambda_item	float	0.1	Weight for the item-side conditional adversarial loss (λ_i).
--fairness_weight	float	0.5	Interpolation weight (α) controlling catalog vs. group parity.

---

📖 Citation

If you use this code or framework in your research, please cite our paper:

@article{FA_BUHS_2026,
  author = {},
  title = {When Popularity Fails: Fair and Equitable Long-Tail Recommendation with Fairness-Constrained Learning},
  journal = {....},
  year = {2026}
}

Note: Please update the citation block with the exact publication details once officially published.

---

📄 License

This project is licensed under the MIT License — see the LICENSE file for details.