Smarter Noise Filtering for Recommendation Systems Using User and Item Reliability

What is it about?

When you browse or click on items online, recommendation systems use that behavior to suggest things you might like. But not every click reflects genuine interest — sometimes people click by accident, or are tricked by misleading thumbnails. This "noise" confuses the system and leads to worse recommendations. Most existing methods try to clean up this noise by looking at each individual click, but they miss a key insight: some users are naturally more random in their behavior, and some items are inherently more likely to attract meaningless clicks. Our method, called EARD, takes a different approach. Instead of only examining each click in isolation, it evaluates the overall reliability of each user and each item based on how they behave across all their interactions. It then combines this "reputation score" with information about individual clicks to decide how much to trust each piece of data during training. The method is lightweight, works with any recommendation model, and only needs two simple settings to tune. Experiments on three large real-world datasets show it significantly outperforms existing approaches — improving recommendation accuracy by up to 27% — while adding almost no extra computational cost.

Featured Image

Photo by Elin Melaas on Unsplash

Photo by Elin Melaas on Unsplash

Why is it important?

This work addresses a fundamental bottleneck in real-world recommendation systems: noisy implicit feedback degrades the quality of suggestions that billions of users receive daily on platforms like e-commerce sites, streaming services, and social media. Unlike previous solutions that are either too computationally expensive for industrial use or require extensive manual tuning, EARD introduces a practical framework that is efficient, easy to deploy, and model-agnostic. Its key innovation — modeling noise at the entity level (users and items) rather than just the interaction level — opens a new direction for building more robust recommender systems. The method's minimal overhead and strong empirical results across diverse settings make it immediately applicable to large-scale production environments where both accuracy and efficiency matter.

Perspectives