Description
Check whether a ranking or recommendation system's outputs are narrowing because users' reactions to them feed back into training, and correct it with randomised exposure or position modelling.
A degenerate feedback loop arises when a system's predictions shape the feedback later used to train it, so items shown more get clicked more and are shown more still. Popularity bias and filter bubbles are the visible symptoms, and the same dynamic can entrench a proxy feature in screening systems. Detection uses diversity measures such as aggregate diversity and average coverage of long-tail items, hit rate computed separately for popularity buckets, and, in production, rising homogeneity of outputs over time. Corrections include giving some exposure at random (as with a random initial traffic pool for each new item), intelligent exploration with contextual bandits, small randomisation analysed with causal inference, and modelling position explicitly, either as a feature or with two models separating the chance an item is seen from the chance it is clicked once seen. Randomisation improves diversity but can harm the experience, so its dose is a design decision.
Attributes
- Discipline
- Computational & Machine Learning
- Loss type
- Siloed
- Stage
- Post-launch
- Data
- Quantitative
- Effort
- Weeks
Steps
- Measure the popularity diversity of outputs (aggregate diversity, long-tail coverage).
- Bucket items by past interaction counts and compare accuracy or hit rate across buckets.
- Track output homogeneity over time in production.
- Introduce limited randomisation or bandit-based exploration to gather less biased feedback, choosing the dose with experience costs in mind.
- Model presentation position explicitly, with positional features or a seen-then-click two-model approach.
- Re-measure diversity and bucket accuracy after the correction.
Sources
- Designing Machine Learning Systems (Huyen, 2022)
Also known as
- Popularity bias audit
- Exposure bias
- Filter bubble detection
- Position bias correction
Categories
- Analytics & Behaviour
- Computational & Machine Learning