What is Position Bias in Recommendation Systems?

Definition
Position Bias is the phenomenon where items shown higher in a list get more clicks not because they are better, but simply because users see them first. This distorts training data and causes models to rank already visible items even higher.
THE PROBLEM
Users scan lists from top to bottom. An item in position 1 might get 30% click through rate while the exact same item in position 5 gets only 5%. If you train a model on this click data, it learns that position 1 items are 6x more relevant. But they are not more relevant; they were just more visible. The model then ranks these items higher, they get more clicks, and the bias amplifies.
WHY IT MATTERS
Without correcting for position bias, your model optimizes for visibility, not quality. Items that happened to be shown first become permanently stuck at the top. New items never get shown, never get clicks, and appear worthless to the model. Over time, recommendations stagnate and user satisfaction drops as the system stops surfacing fresh content.
THE FEEDBACK LOOP
Position bias creates a self reinforcing cycle: item ranked high gets clicks, model sees clicks as relevance signal, item gets ranked even higher, gets more clicks. This is a feedback loop. Without intervention, the system converges to showing the same small set of items regardless of actual user preferences.
💡 Key Insight: Position bias is not just noise in your data. It actively corrupts your model and gets worse over time unless explicitly corrected.

💡 Key Takeaways

✓Items in position 1 get up to 6x more clicks than identical items in position 5, purely from visibility

✓Training on raw clicks teaches models that visibility equals relevance - fundamentally wrong

✓Feedback loops amplify bias: high rank → more clicks → higher rank → even more clicks

✓Without correction, recommendations stagnate and new items never surface

✓Position bias gets worse over time as the model reinforces its own mistakes

📌 Interview Tips

1Explain the visibility trap: 30% CTR at position 1 vs 5% at position 5 for identical items

2Describe the stagnation symptom: same items dominate top positions for weeks

3Discuss how to detect: shuffle positions randomly for 1% of traffic and compare CTR curves

← Back to Position Bias & Feedback Loops Overview