Fairness in PCA-Based Recommenders

Understanding Recommender System Unfairness

• 💡 PCA (Principal Component Analysis) is a common dimensionality reduction technique used in recommender systems, but it can inadvertently lead to unfairness.
• 🎯 Unfairness often arises from modeling choices and data imbalances, not malicious intent, particularly affecting minority and niche user groups.
• 🧠 Systems tend to perform better for mainstream content due to more available data, while niche users are harder to characterize and learn from.

PCA's Limitations in Collaborative Filtering

• ⚠️ PCA aims to find the best overall approximation of an interaction matrix, which can cause it to neglect subgroups or niche items where data is less dense.
• 📚 Research in "When Collaborative Filtering Is Not Collaborative" shows PCA's focus on approximation can lead to over-specialization on popular content.
• 📉 This over-specialization means PCA can under-represent niche items and even prevent popular artists from being recommended to new potential fans, effectively "memorizing" existing user bases.

Addressing Unfairness with Item-Weighted PCA

• 🛠️ A proposed solution is item-weighted PCA, which involves intelligently boosting less popular artists to ensure their group information (e.g., genres) is captured.
• ✅ This approach uses a tunable parameter to control the level of boosting, aiming to reduce "diagonal entries" in the similarity matrix that indicate over-specialization.
• 🚀 Item-weighted PCA has shown it can improve both fairness and performance simultaneously, challenging the common assumption of a trade-off between these goals.

The Role of Power Niche Users

• 🔑 Power niche users are highly active individuals with specialized interests who generate valuable data for the platform.
• 📈 Upweighting these specific users in the loss function can help filter out noisy, low-popularity data and focus on valuable "hidden gems."
• 📊 This strategy recognizes that good embeddings, which accurately capture diverse user interests, ultimately benefit all users and improve overall system performance.