Skip to content

CONCEPT Cited by 1 source

Click-duration reweighting

Definition

Click-duration reweighting is a training-data preprocessing technique that attenuates noisy binary click labels by scaling their training weight with a monotonic function of click dwell time. A click with 30 seconds of dwell contributes more gradient than a click with 1 second of dwell; a sub-second bounce click contributes almost nothing.

Canonical shape (Source: sources/2026-04-27-pinterest-from-clicks-to-conversions-architecting-shopping-conversion-candidate-generation):

w = f(log(1 + t / t_max))

where t is the non-negative click duration in seconds and t_max is a tunable cap. The logarithmic shape compresses the long tail of very-long dwells (watch-till-end, still-on-page) so a 10-minute dwell doesn't swamp a batch's loss, and the t_max cap prevents overweighting pathological outliers.

Why it's used

Click is a cheap, abundant engagement signal but it's noisy:

  • Bounce clicks — accidental taps, instant back-outs, pagination misfires. Binary-coded as positives, but the user's intent is clearly not engagement.
  • Content-mismatch clicks — the user clicked, saw the landing page was wrong, left immediately. Technically a click, but training on these teaches the model to serve low-quality landing pages.
  • Positive-only masking: without reweighting, the model learns that any click is a positive signal of equal value, which systematically biases training toward content that extracts clicks but not engagement.

Click duration is a cheap proxy for engagement depth. Short dwell ≈ noisy / accidental click; long dwell ≈ genuine engagement. Reweighting by dwell converts a noisy binary label into a continuous-weight positive that better reflects latent user intent.

The Pinterest shape

Pinterest uses click-duration reweighting in the context of dual positive signals for the shopping conversion candidate generation model:

"We supplement primary conversion signals with onsite engagement data (clicks, repins). This broadens data coverage, improving model generalization and ad funnel survival rates. To mitigate click data noise and decrease false positive clicks, we apply a log-based re-weighting function w based on the click duration: [...] where t is the non-negative click duration in seconds and t_max is a tunable constant used to cap the re-weighting function."

The reweighting is applied at the positive-contribution level — a longer click contributes a larger weight to the positive term in the contrastive loss. A 0-second click contributes weight 0; t_max-second-and-beyond clicks saturate at the cap.

Shape of the function

f(log(1 + t / t_max)) has three desirable properties:

  1. Monotonic in t — longer dwell ⇒ larger weight. Matches the latent-intent gradient.
  2. Concave (log-shaped) — diminishing returns for very long dwells. A 60-second dwell isn't 60× more valuable than a 1-second dwell; the log-scaling keeps the contribution sensible.
  3. Saturates at t_max — prevents a handful of extreme-dwell outliers from dominating gradient updates.

The exact outer function f and the value of t_max are not disclosed in the Pinterest post; both are tunable surfaces.

When to apply

Click-duration reweighting is applicable when:

  • The training signal is click-based but the downstream objective is deeper-than-click (engagement depth, conversion, revenue).
  • Click dwell time is observable and trustworthy — requires browser/client-side instrumentation that captures dwell before the user navigates away.
  • Per-click variance in dwell time is high — if all clicks are roughly the same duration, reweighting adds no signal.

When NOT to apply

  • When dwell time is not observable (offline batch signals, external conversion-only pipelines).
  • When dwell time is not correlated with the downstream target — some conversion funnels have short dwell (rapid add-to-cart on familiar products) as a strong signal.
  • When bot traffic could produce long-dwell anomalies (automated scrolling, hanging connections).

Adjacent techniques

  • Minimum-dwell filtering — binary variant: drop all clicks with dwell < threshold. Simpler, cruder; loses the gradient information for mid-dwell clicks.
  • Implicit-feedback weighting (Hu, Koren, Volinsky 2008, original ALS confidence-weighting paper) — similar idea in CF recommendation: engagement intensity as a confidence weight.
  • Watch-time weighting (YouTube) — equivalent shape for video: watch duration reweights a watch-started label.
  • Dual positive signal — at Pinterest, click-duration reweighting is the noise-control layer inside the dual-positive setup.

Caveats

  • Dwell is gameable at the adversary level — advertisers can craft landing pages that trap users (intentional friction, long-load pages) to inflate dwell. Adversarial robustness is not discussed in the Pinterest post.
  • Dwell requires foreground page instrumentation — backgrounded tabs, app-switching, or device-locked states can produce misleading long dwells.
  • Post-click behaviour varies by surface — a Search click has different dwell distribution than a Home Feed click. Surface-specific t_max may be needed.
  • No absolute weight numbers disclosed. Pinterest names the functional form but not t_max, per-click weight range, or effect-size compared to binary clicks.

Seen in

Last updated · 445 distilled / 1,275 read