Feedback loops and exposure bias

A music recommender that promotes already-popular artists gives them more opportunities to receive plays. Those additional plays enter the next training dataset and may reinforce the same ranking, even when other artists would also have interested listeners. The observed outcome therefore reflects both user preference and the opportunities created by the earlier ranking policy.

The same selection effect appears when moderation sends only flagged items for review or when a fraud model blocks payments before an outcome can occur. Some parts of the task then receive more reliable labels than others. Training on those observations without accounting for how they were selected can preserve the system’s existing blind spots.

Log exposures, positions, candidate sources, and policy or experiment versions. If a controlled exploration policy selects actions probabilistically, retain the actual selection probabilities needed by the intended estimator. A predicted click score is not the same as an action-selection probability.

Carefully bounded exploration can create evidence about underexposed items, but it must respect safety, eligibility, and user experience. Independent audits or human judgments may be more appropriate in some domains. There is no obligation to expose harmful or unauthorized content in order to obtain labels.

A propensity is the probability that the logging policy selected an action in its observed context. Inverse-propensity weighting gives more weight to outcomes from actions that were less likely to be selected, using those probabilities to account for differences in exposure. It requires reliable probabilities and coverage of the actions being evaluated, because an action that was never selected has no observed outcome to reweight. Very small probabilities can also make the estimate unstable.

Position effects complicate click labels even after an item is shown. A user may click the first acceptable result rather than compare every option. Separate position, exposure, and relevance as far as the data allows, and validate any correction against independent evidence or a controlled experiment.

Before choosing a correction method, establish what the system actually logged and which actions had a chance to occur. That explanation connects the statistical assumption to the design of data collection. Where the required information is missing, an audit or a carefully controlled experiment may be needed before the effect can be estimated credibly.

Track catalog coverage, new-item exposure, repeated recommendations, and user or creator segments alongside immediate engagement. A short-term improvement can reduce variety or make it harder for new content to accumulate evidence. Define these tradeoffs with the product rather than assuming engagement captures everything.

Longer-running experiments or suitable holdout groups can help distinguish a lasting change in user experience from a short-term shift in exposure. Recording how the logging policy changed keeps those findings connected to the training data. This completes the connection from requirements through operation and back to new training examples, which you can apply in the practical designs. The final topic explains how to present and discuss those decisions in an interview.