Interleaving Experiments for Ranking Models

Interleaving is an online evaluation technique for comparing ranking systems. Rather than assigning different users or sessions to separate rankers, it combines results from two or more rankers into a single list shown for a query. The method tracks which ranker contributed each item. User interactions such as clicks can then provide pairwise evidence about which ranking better served that shared context.

In a hypothetical comparison, one ranker returns a, b, c and another returns b, d, a. An interleaving procedure chooses items from each list, handles duplicates and records attribution. If users click more items attributed to one ranker across many comparisons, that can suggest a preference under the experiment's setup. Specific methods such as team-draft or probabilistic interleaving differ in how items are selected and how credit is assigned.

Because both rankers share queries and users, interleaving can reduce variance from differences in query difficulty or audience composition. It may require less traffic than a conventional A/B test for detecting some pairwise ranking preferences, but results are not universally interchangeable with A/B outcomes. The displayed list itself is a mixture, so it may not match the experience of either complete ranking. Position bias, duplicate items, click propensity, trust and novelty effects can distort attribution.

Interleaving is most useful as a comparison tool for ranking systems under a defined interaction signal. It does not directly measure every product outcome, such as retention, revenue, accessibility or latency. Define query eligibility, experiment duration, attribution method and statistical analysis in advance. Use separate evaluation for broad release decisions, particularly if ranking changes affect safety or exposure. Interleaving can efficiently identify promising candidates, while A/B or controlled rollout testing assesses the complete experience and guardrails. Interpret evidence as preference under the chosen protocol, not proof of a universal ranking winner.