Skip to content

CONCEPT Cited by 1 source

Soft-spacing penalty

Definition

A soft-spacing penalty is a distance-weighted score reduction applied to a candidate item when nearby already-placed items share a sensitive content class. Graceful alternative to hard filtering: the candidate can still win the slot if its other scores are high enough, but clustered same-class content is actively discouraged.

Formalised by Pinterest (Source: sources/2026-04-07-pinterest-evolution-of-multi-objective-optimization-at-pinterest-home) for Home Feed's Blender:

qᵢ(t) = 𝟙[cᵢ ∈ R] · Σ_{d=1}^w (1/d) · 𝟙[c_{t−d} ∈ R]
  • cᵢ — content class of candidate i at decision position t.
  • R — the sensitive set of classes requiring spacing (e.g. content with elevated quality risk).
  • w — backward window size.
  • 1/d — inverse-distance weighting: a same-class item one slot back contributes more than one w slots back.
  • Indicator 𝟙[·] — penalty only activates when both the candidate AND a nearby placed item are in R.

Integrated into the SSD utility:

U'ᵢ(t) = Uᵢ(t) − λ · qᵢ(t)

where λ balances soft-spacing against engagement and diversification objectives.

Why soft-spacing over hard filtering

Pinterest's explicit framing:

"This is an important next step for improving content quality on Pinterest and protecting users from content that warrants additional caution, where in the past we usually rely on strong enforcement like filtering which sometimes leads to less satisfying user experience if there is no backfill."

Hard filtering fails when:

  • Backfill is insufficient — removing a class entirely can leave feed gaps.
  • Binary decisions ignore gradient — slightly-elevated-risk content shouldn't be treated the same as actively-harmful content.
  • User experience suffers — visible holes or abrupt topic shifts look broken.

Soft-spacing threads the needle: reduce exposure clustering without removing content; preserve candidate pool size; keep the slate flowing.

Rollout at Pinterest

  • Mid 2025 — soft-spacing penalty launched on content with elevated quality risk.
  • Late 2025 — abstracted via config-based framework to extend to more sensitive classes as needed.

Design parameters

  • Sensitive set R — which content classes require spacing. Per-class configurable.
  • Window size w — how far back the penalty considers. Too small = tight clusters still allowed; too large = excessive dispersion.
  • Distance kernel — inverse 1/d in Pinterest's instance; exponential decay, piecewise-constant, or other kernels plausible.
  • Weight λ — trades soft-spacing against other objectives. Per-class tunable.

Caveats

  • Not a substitute for policy enforcement. Actively-harmful content should still be filtered or removed upstream; soft-spacing handles the gradient of borderline and elevated-risk content.
  • Measurement is tricky — soft-spacing affects slate composition more than item-level metrics. Evaluation needs slate-level (feed-level) measurement and multi-week soak to surface long-term effects.
  • Penalty-coefficient tuning — Pinterest doesn't disclose λ values; poorly tuned penalties can over-suppress legitimate content.
  • Composes with diversification objectives — the soft-spacing term adds to SSD's spectral penalty; interaction effects aren't discussed in the source.
  • Class-definition upstream — requires classifiers upstream that produce the cᵢ signal; quality of that classification bounds the penalty's utility.

Seen in

Last updated · 319 distilled / 1,201 read