Problem
Search and recommendation lists often show only some groups of people, for example one skin tone range for fashion and beauty requests. A ranker that sorts by utility alone reproduces what the data favors, so a group can vanish from the top positions. Reordering cannot fix this when the retrieved candidates contain no items from that group. Diversity must be added without losing engagement or adding much latency.
Example from the article: Pinterest Related Products for the Pin "Shirt Tail Button Down" before and after diversification.
Solution in brief
Make both stages diversity-aware. At ranking, rebuild the utility-sorted list with Round Robin over groups or a Determinantal Point Process (DPP). At retrieval, give each group enough candidates by overfetching, bucketized nearest-neighbor aggregation or the Strong-OR operator. Trigger it only for relevant categories and judge it by A/B tests.
Questions this case answers
- How to guarantee that every group of a chosen attribute appears in the top results of a multi-stage retrieval and ranking system without hurting engagement
- How to make a candidate generation stage return enough items from each group when a downstream reranker can only reorder what it receives
- How to make search and recommendation results show every group of a chosen dimension, such as skin tone range, in the top positions?
- How to diversify results when the retrieved candidate set lacks items from some groups?
Tools and methods
- Round Robin (RR) reranking — take the top item of each group in turn.
- DPP reranking — list-wise utility versus similarity trade-off.
- Overfetch-and-Rerank — fetch more candidates until each group has a minimum.
- Bucketized ANN retrieval — keep top candidates per group in aggregation.
- Strong-OR operator — search operator with a minimum share per criterion.
- Triggering logic and A/B tests — apply it only where it fits and track guardrails.
Google or a one-time email code. No password.