Breaking the Loop: An Empirical Comparison of Strategies for Novelty and Freshness in YouTube Music

Modern music recommender systems rely on continuous training loops where models are frequently updated based on recent user engagement. While this ensures the system stays relevant, it creates a feedback loop that favors familiar, previously consumed content. This 'popularity bias' suppresses two critical categories: new releases (temporal freshness) and unlistened catalog items (novelty). This paper evaluates various strategies to break this loop, specifically examining how interventions at different layers of the recommendation stack—serving, training data, architecture, and exploration—impact discovery and user engagement on the YouTube Music homepage.

The researchers conducted off-policy online A/B tests for six distinct interventions, keeping upstream candidate generation fixed to isolate the effects of each strategy. The interventions were categorized into four layers:

• Serving: Applying heuristic boosts or diversification rules directly to the ranked list.
• Training Data: Reweighting recent engagement data to prioritize new content.
• Architecture: Using auxiliary 'bias towers' to disentangle presentation bias (e.g., position) from true user preference.
• Exploration: Utilizing uncertainty estimates from a Spectral-normalized Neural Gaussian Process (SNGP) to either boost scores at serving or prioritize high-uncertainty examples during training.

The study reveals that the choice of intervention layer significantly dictates the outcome. Serving-time heuristics, while simple to implement, are quickly 'learned away' by the model as it continues to train on the resulting biased feedback. Architectural debiasing successfully reduces popularity dominance and improves diversity metrics but fails to drive meaningful discovery of new content, often introducing hidden integration costs. The most successful approach for surfacing new releases was uncertainty-driven exploration using the SNGP head, which explicitly directs the system to items it knows less about. However, the authors emphasize that these gains come with a necessary tradeoff in short-term engagement or diversity, highlighting that breaking a feedback loop requires sacrificing immediate precision for long-term system health.