EngineersofAI LogoCoursesBlogResearch LabAI LettersThe LabCode BankInteractive 3DKodrEarnestJobs
Skip to main content
Interactive 3D/Ad Click-Through Rate (CTR) Prediction
Predicted CTR
2.300%
Bid
$5.00
eCPM
$115.00
Model AUC
0.81
CTR Prediction Pipeline
Ad Request
User page load / search query
Feature Assembly
User + Ad + Context features
CTR Model: Wide & Deep
AUC 0.81 | Latency 8ms (GPU)
CTR Score
Predicted CTR: 2.30%
Auction
eCPM = CTR × Bid = $115.00
Ad Shown
Highest eCPM wins the auction
User Features
· Age, gender, location
· Interest categories
· Historical CTR
· Device type, OS
· Recent search queries
Ad Features
· Ad category / topic
· Creative format
· Advertiser history
· Bid amount
· Landing page quality
Context Features
· Time of day / day of week
· Page/query context
· Device & browser
· User session depth
· Geographic context
ModelAUCLatencyGPUScale
Logistic Regression0.740.5msNoBillions/day
Wide & Deep0.818msYesBillions/day
DLRM0.8415msYesBillions/day
Online learning: model embeddings updated every hour from new click/no-click feedback (FTRL - Follow-The-Regularized-Leader). Sparse feature IDs hashed into fixed embedding table. Cold-start new ads get exploration budget before exploitation.
Controls
Model Architecture
User Segment
Moderate history. User embeddings partially trained. CTR estimate has moderate confidence.
Advertiser Bid (CPM $)$5
$0.50$50
Auction: eCPM = predicted CTR × max bid. Not highest bidder wins - best product of relevance and bid.

DLRM: dense features → bottom MLP → embeddings → feature interaction layer (dot products) → top MLP → CTR.

Scale: Google runs ~8.5 billion ad auctions per day. Latency SLA: <20ms total auction.

Ad Click-Through Rate (CTR) Prediction - Interactive Visualization

Ad click-through rate prediction is the highest-value ML problem in industry, directly governing advertising revenue at Google, Meta, and every ad-supported platform. The system assembles user features (demographics, interests, historical CTR), ad features (creative, category, bid), and context features (time, device, page) into a single feature vector. Logistic Regression remains a strong baseline due to its speed and interpretability. Wide and Deep networks (Google, 2016) memorize frequent patterns via the wide component while generalizing to unseen user-item pairs via the deep component. DLRM (Facebook, 2019) uses dedicated MLPs for dense and sparse features with learned pairwise feature interactions. The auction multiplies predicted CTR by the advertiser's bid to compute eCPM - the highest eCPM wins, not the highest bid.

  • eCPM = CTR × Bid × 1000: relevance and bid together determine auction winner, not bid alone
  • Wide & Deep: wide component memorizes user-item-context co-occurrences; deep component generalizes via embeddings
  • DLRM: separate bottom MLP for dense features, embedding lookup for sparse IDs, dot-product interaction layer
  • Cold start: new users with no history fall back to demographic + contextual features - CTR estimates are unreliable
  • Online learning: FTRL (Follow-The-Regularized-Leader) updates embeddings every hour from new click feedback
  • Scale: Google processes ~8.5 billion auction requests per day with a total latency SLA under 20ms

Part of the EngineersOfAI Interactive 3D - free interactive visualizations covering every major concept in machine learning and AI engineering. Hover any element for a plain-English explanation. No code required.