Skip to content

CONCEPT Cited by 1 source

Feed-level reranking

Definition

Feed-level reranking (a.k.a. slate reranking, feed-composition reranking) is a final reranking stage whose objective is what composition of items to serve, not which items are individually best. Operates over a ranked candidate set produced by upstream pointwise ranking and produces the actual served slate by optimising slate-level properties (diversity, topic balance, quality spacing, business constraints).

Canonical contrast (Source: sources/2026-04-07-pinterest-evolution-of-multi-objective-optimization-at-pinterest-home):

"While earlier stages mostly optimize for certain positive actions (e.g., saves) given an impression, the multi-objective optimization layer tackles a different problem: determining the best composition of a feed served to the user."

Why it's a distinct funnel stage

Pointwise rankers optimise per-candidate engagement likelihood. That objective does not see the slate:

  • Two near-identical candidates both score high — pointwise ranker produces [A, A', A''...].
  • User session prefers variety — the served feed collapses engagement signals (less clicking, shorter session, lower revisit).
  • Closed feedback loop — subsequent training data is shaped by this collapse; the ranker reinforces the mono-feed bias (exposure-bias-style dynamic).

Feed-level reranking inverts the optimisation axis: composition matters. A diverse slate may have lower mean ranker score but higher slate utility.

Relation to retrieval → ranking funnel

Extends the canonical retrieval → ranking funnel with a third production stage:

retrieval  →  pre-ranking  →  ranking  →  multi-objective feed-level reranking
(10⁶→10⁴)     (10⁴→10³)       (10³→10²)    (10²→~30 slots)
pointwise    pointwise         pointwise      SLATE-LEVEL
recall      engagement        engagement     composition

The final stage operates over a small slate (tens of items) so slate-level algorithms (DPP, SSD, soft-spacing) that would be intractable at earlier stages become affordable.

Common algorithms

  • DPP — slate-global optimisation via a kernel matrix combining relevance + similarity.
  • SSD — position-adaptive windowed spectral decomposition.
  • Soft-spacing penalties — class-based distance-weighted dispersion, composable into an existing utility equation.
  • Legacy heuristics — fixed category gaps, MMR (Maximal Marginal Relevance), hand-tuned spacing rules.
  • Forthcoming — Pinterest names "unified generative post-ranking model" and "reinforcement learning based value model" as active work.

Why Home Feed particularly needs this

Pinterest's observation: "users tend to have lower intent when visiting Home Feed and their browsing behavior will be significantly impacted by what they see. For example, visually repetitive content is less engaging and is likely to reduce the user's session length and the likelihood that a user will revisit Pinterest."

Compare to intent-heavy surfaces (Search, Related Pins) where users arrive with a query or context — there, composition matters less because the query itself constrains variety.

Operational implications

  • Multi-week evaluation — feed-level effects show up on session-length and retention, not CTR. Short A/B tests miss the signal.
  • Slate-level metrics — need to measure diversity (e.g., pairwise similarity distribution, category entropy) alongside engagement.
  • Feedback-loop guardrails — a weak reranking stage lets the upstream ranker's pointwise bias dominate, which trains subsequent rankers on less-diverse impression distributions.

Caveats

  • Not a substitute for good pointwise ranking — feed-level reranking can rebalance, not generate. If the ranker surfaces bad candidates, reranking can't fix it.
  • Business-constraint composition — real production rerankers also encode revenue, supply-side, policy, and ad-load constraints; the academic "diversity" framing is a subset.
  • Latency budget — slate reranking typically gets tens of milliseconds; slate-global optimisation (DPP) can eat that budget, pushing production toward linear-time approaches like SSD.
  • Candidate-set dependence — slate reranking quality is bounded by the candidate set's own variety; too-small or too-biased upstream outputs constrain what the reranker can produce.

Seen in

Last updated · 319 distilled / 1,201 read