Unheard Voices 🎵

Investigating Gender Bias in Music Recommendation Systems

📌 Project Overview

This project investigates the presence and mitigation of gender bias in music recommender systems. We explore two real-world datasets (Spotify Million Playlist Dataset and Last.fm), assess fairness in popularity-based recommendation outputs, and apply algorithmic debiasing using AIF360 and Fairlearn.

🎯 Project Goals

• Analyze the gender distribution in Spotify and Last.fm music data.
• Build baseline recommenders based on item popularity.
• Quantitatively evaluate fairness using statistical metrics (e.g., disparate impact).
• Apply bias mitigation strategies (e.g., Reweighing, Fairlearn reductions).
• Promote fairer exposure for underrepresented gender groups in algorithmic music curation.

❓ Research Questions

• Does gender bias exist in music recommender systems?

Sub-Research Questions:

• How do different fairness metrics e.g. demographic parity, equal opportunity reflect imbalances?​
• How do different fairness-aware tools perform in mitigating gender bias in music recommendation systems

🧪 Methodology

We followed the Design Science Research approach (Peffers et al., 2007):

1. Problem Identification – Gender bias in music recommender systems.
2. Design & Development – Baseline + Content-Based recommenders using Spotify & Last.fm.
3. Evaluation – Fairness and accuracy metrics (exposure ratio, disparate impact, etc.).
4. Tools Used: Fairlearn, AIF360, FaiRecSys, RecBole.

📁 Structure

• data/: contains the datasets used or links to download them
• notebooks/: contains data processing, descriptive statistics, modeling, and fairness evaluation notebooks
• requirements: packages needed to run the notebooks
• License: open-source license for usage and citation
• README.md: this file

📊 Datasets

• Spotify: A dataset of 1 million user-generated playlists (2010–2017) used to analyze track popularity and gender representation in music recommendations Million Playlist Dataset
• Last.fm: Collected via publicly available sources from Kaggle
• The datasets were enriched using the MusicBrainz API, by retrieving the gender attributes

Some data files (due to size or licensing) may be partially included or require external download.

🧠 Tools & Libraries

• Python (pandas, numpy, scikit-learn)
• Fairlearn
• AIF360
• Jupyter Notebooks

📈 Results Summary

✅ Gender bias detected in both baseline and content-based recommenders. ⚙️ Fairness-aware re-ranking methods reduced bias while preserving recommendation quality. 🎯 AIF360 and Fairlearn metrics guided bias mitigation strategies effectively.

👩‍💻 Contributors

• Patricia Haumer (h11910653)
• Katharina Rosa Pöcher (h11917060)