Skip to content

CONCEPT Cited by 2 sources

CTR prediction

Definition

CTR prediction (click-through rate prediction) is the machine-learning task of estimating the probability that a user will engage (click, like, watch, tap) with a specific candidate item (ad, content, link, Pin) in a specific context. It is the core scoring primitive under ads ranking and recommendation systems: given user, candidate, and context features, output a probability P(engage | user, candidate, context).

CTR prediction is typically trained on binary engagement labels (engaged / not engaged) at the impression-level, with log-loss or a calibrated classification objective. In multi-task / multi-label recsys architectures (see MTML ranking), CTR is often one of several task heads (CTR + like + share + follow, each with its own head) trained jointly.

Why it's load-bearing

The CTR score enters the final ranking formula and drives the auction mechanics in ads serving:

  • Ad ranking. Candidate ads are ranked by bid × predicted-CTR (or a more complex expected-value formula); higher-CTR ads win more auctions.
  • Calibration matters for the auction to be economically valid. Predicted probabilities must be close to empirical rates — overconfident CTR predictions bias budget spend to specific surfaces or ad formats and break revenue guarantees. See surface-specific calibration.
  • Serving cost is multiplied by candidate count. Every ad candidate needs a CTR score, so CTR models are evaluated O(candidates) times per request — the scaling axis behind optimisations like request-level user-embedding broadcasting and surface-specific tower trees.

Canonical wiki instances

Pinterest unified ads engagement model — surface-specific calibration

Pinterest's unified ads engagement model predicts CTR across three ads surfaces (Home Feed, Search, Related Pins) with surface-specific calibration as the critical correctness mechanism (Source: sources/2026-03-03-pinterest-unifying-ads-engagement-modeling-across-pinterest-surfaces):

"Since the unified model serves both HF and SR traffic, calibration is critical for CTR prediction. We found that a single global calibration layer could be suboptimal because it implicitly mixes traffic distributions across surfaces."

Instacart Carrot Ads — Wide-and-Deep pCTR with Domain Adaptive Learning

Instacart's Carrot Ads omnichannel retail-media platform runs CTR prediction across many partner retailer sites under a multi-tenant topology where each partner has its own data distribution. The partner-specific pCTR model is a Wide-and-Deep architecture trained with Domain Adaptive Learning — Instacart Marketplace as source domain, partner site as target domain (Source: sources/2026-05-04-instacart-empowering-carrot-ads-with-domain-adaptive-learning).

Quote on the model:

"This model predicts CTR by first transforming raw inputs, like user IDs and product text, into dense feature embeddings. These features are concatenated and processed through two parallel paths: an interaction layer for learning explicit feature interactions and a deep Multi-layer Perceptron (MLP) tower for learning complex, hidden patterns. ... Finally, a Sigmoid activation squashes the result into a probability score (pCTR) between 0 and 1. This architecture combines a linear 'wide' model (for memorization of specific feature interactions) with a 'deep' neural network (for generalization)."

Distinctive properties of this CTR-prediction instance:

  • Multi-tenant: a single pCTR-prediction stack serves many partner sites, each with cold-start exposure (new-partner cold-start).
  • Cross-domain transfer: shared shopping-context embeddings + per-partner fine-tune (patterns/cross-domain-warm-start-via-shared-embeddings).
  • Real-time auction latency: the model runs in a real-time partner ad auction; per-partner feature trimming keeps the model lightweight.
  • First-party-data moat: the source-domain Marketplace data carries signal partners structurally cannot replicate — the reason DAL outperforms from-scratch training even when the target has sufficient data.

Seen in

Last updated · 542 distilled / 1,571 read