Preranking

Preranking is a stage in information retrieval and search systems in which a subset of items retrieved by an initial candidate generation step is scored and often reordered using a lightweight model or heuristic before a final, more expensive ranking stage. The goal is to reduce latency and computational cost while preserving high relevance.

In a typical pipeline, candidate generation retrieves a large set of candidate documents, products, or answers. Preranking then filters and reorders this set using fast models to produce a smaller list that is fed to the final ranking component, which applies more sophisticated, resource-intensive scoring. Preranking can also adjust ordering with simple rules or feature-based scores, integrating signals such as query-document features, user context, and historical interactions.

Common approaches include linear models, logistic regression, and gradient-boosted trees trained on lightweight feature sets, and sometimes shallow neural networks. Features may include lexical matches, basic document statistics, and user or session features. Additional techniques involve approximate nearest neighbor search or rule-based pruning to quickly discard unlikely items. The architecture is often described as a cascaded ranking pipeline.

Evaluation focuses on a trade-off between latency, throughput, and retrieval quality. Offline metrics such as precision, recall, and NDCG are used, along with online experiments like A/B tests to ensure the preranking step does not degrade end-to-end performance. Preranking must avoid discarding relevant results and handle data drift and feature staleness.

Domains commonly employing preranking include web search, e-commerce recommender systems, and question-answering or conversational systems. The term is sometimes used interchangeably with coarse ranking or candidate generation improvements; in practice, preranking is an essential component of many modern cascaded ranking architectures.