Interleaving

Definition

Interleaving merges two ranked result lists into a single ranking by alternating (“zip-merge”) items from each, optionally weighted by sampling probabilities so one source contributes more often. It serves two purposes:

  1. Fusion baseline — a simple, hard-to-beat way to combine multiple retrievers (e.g. lexical + vector) when their scores are non-comparable, without training a model.
  2. Online evaluation — interleaving two rankers and observing which side’s items get clicked is a sensitive B testing alternative that needs less traffic than split testing.

Why it matters

In a multi-retriever / Hybrid Search setup, interleaving solves the cold-start problem: you need behavioral history before you can train an LTR fusion model, but you need some unified ranking to collect clicks. Interleaving provides that initial ranking, after which a LambdaMART / LTR re-ranker (e.g. Metarank) can take over.

Interleaving as online evaluation

The second use case deserves its own treatment, because it solves a problem A/B testing cannot.

Why parallel ranking A/B tests break. Simultaneously running ranking experiments violate the independence assumption — they influence each other’s outcomes — so you cannot simply run many at once. The workarounds (multivariate experiments, splitting traffic into subgroups with one experiment each, or leaning on indirect early-indicator metrics) each cost either statistical power or considerable complexity.

Team-draft interleaving. Instead of randomizing on the customer, results from two rankers are merged into one list and each position becomes the unit of randomization. You then measure whether users interact more with results attributed to ranker A or ranker B, using bootstrapping to obtain a distribution over the relative wins.

Why it is dramatically more sensitive. The clearest explanation comes from segmenting users by purchase intent:

SegmentBehaviorContribution to an RCT
Low intentBrowse, rarely convertNoise — searches and pageviews, no conversions
High intentConvert almost regardless of rankingNoise — they convert even on a bad ordering, so they carry no evidence about quality
ConvincibleConvert only if shown the right itemThe only real signal — and extremely sparse

An RCT learns almost solely from that last sliver. Interleaving also extracts signal from high-intent users, because even a certain-to-convert user still picks the item that fits best — and that pick is attributable to one ranker. The number of meaningful votes rises substantially.

Booking.com reported potential for a 10–100x speedup from this sensitivity, allowing tests on less traffic with more variants in parallel. On statistics alone their runtime could drop to hours, though they still run at least a day so no decision rests on a single timezone’s users.

Limitations — interleaving is a preselection tool. A conclusive preference in an interleaving trial does not imply the corresponding metric will move significantly in an RCT at the usual expected effect size, and interleaving cannot estimate effect size at all. So it narrows the field of candidate rankers, and a confirmatory RCT still follows. The worst case is promoting a suboptimal model into that RCT rather than the best one.

Articles