Feedback Loops: How Bias Amplifies Over Time

AMPLIFICATION DYNAMICS
Feedback loops occur when model predictions influence future training data. Show item A at position 1, it gets clicks, model learns A is good, shows A at position 1 again, more clicks. After a few retraining cycles, item A dominates regardless of actual quality. Items never shown become invisible: no exposure means no clicks means no data to learn relevance.
CONVERGENCE TO POPULARITY
Without intervention, recommendation systems converge to showing popular items to everyone. The rich get richer: popular items get more exposure, more clicks, appear more relevant, get even more exposure. Niche items that would delight specific users never surface. This hurts long tail discovery, reduces diversity, and eventually reduces overall engagement as recommendations become generic.
BREAKING THE LOOP
Three strategies to break feedback loops: exploration (randomly show some items to gather unbiased data), counterfactual training (weight training examples by how surprising they were), and holdout groups (reserve some users who never see personalized recommendations to measure true baseline).
💡 Key Insight: A recommendation system without exploration is optimizing for yesterday preferences. It cannot discover that users would love something they have never been shown.
MEASURING LOOP SEVERITY
Track diversity metrics over time. If catalog coverage (percentage of items shown to at least one user) drops from 80% to 40% over months, you have a severe feedback loop. If average item age keeps increasing, new items are not surfacing. If user engagement plateaus despite growing catalog, the system is stuck.

💡 Key Takeaways

✓Model predictions influence training data: shown items get clicks, unseen items get nothing

✓Rich get richer: popular items get more exposure → more clicks → appear more relevant

✓Without intervention, systems converge to generic recommendations that hurt long-tail discovery

✓Break loops with exploration (random items), counterfactual training, and holdout groups

✓Monitor catalog coverage: dropping from 80% to 40% indicates severe feedback loop

📌 Interview Tips

1Describe the amplification: item A shown at pos 1 → clicks → model ranks higher → more clicks

2Explain diversity metrics: catalog coverage, average item age, engagement plateau

3Discuss holdout groups: 1-5% of users see non-personalized results to measure baseline

← Back to Position Bias & Feedback Loops Overview