CTR Prediction
“What is the click probability for this ad impression?”
Book a demo and get a free trial of the full platform: research agent, fine-tune capabilities, and forward-deployed engineer support.
By submitting, you accept the Terms and Privacy Policy.
Loved by data scientists, ML engineers & CXOs at
A real-world example
What is the click probability for this ad impression?
Traditional CTR models rely on user-level features and ad metadata, missing the cross-entity signals that actually drive clicks: which publishers attract which user segments, how creative fatigue propagates across campaigns, and which ad-user pairings historically convert. For an ad platform serving 10B impressions per day, a 5% CTR lift translates to $120M in additional annual revenue.
Quick answer
Graph neural networks predict click-through rates by learning cross-entity patterns between users, ads, publishers, and campaigns. Instead of hand-engineering features for months, a GNN reads the relational structure directly and discovers signals like 'users who clicked similar creatives on related publishers,' producing 5-15% CTR lift over flat-table baselines.
Approaches compared
4 ways to solve this problem
1. Logistic regression / feature-engineered models
Build user-level and ad-level features (demographics, ad category, device type), then train a logistic regression or shallow model on click/no-click labels.
Best for
Fast baseline when you have limited data and a small engineering team. Easy to explain and audit.
Watch out for
Requires months of feature engineering. Misses cross-entity interactions like publisher-audience affinity and creative fatigue propagation across campaigns.
2. Deep learning (wide-and-deep, DeepFM)
Combine memorization (wide component) with generalization (deep component) to model feature interactions. Popular at Google, Huawei, and other ad platforms.
Best for
Good at learning feature crosses from sparse categorical data. Strong when you have billions of training examples.
Watch out for
Still operates on a flat feature table. Cannot capture multi-hop relationships like 'this user's segment historically converts on this publisher's inventory for this advertiser category.'
3. Gradient boosted trees (XGBoost, LightGBM)
Train an ensemble of decision trees on tabular features. The go-to baseline for most ad tech prediction tasks.
Best for
Strong out-of-the-box performance on structured data. Handles missing values and mixed feature types well.
Watch out for
Treats each impression independently. Cannot learn from the graph structure connecting users to ads to publishers to campaigns without extensive manual feature joins.
4. KumoRFM (relational graph ML)
Point Kumo at your raw impression, click, campaign, and publisher tables. Write a two-line PQL query. The GNN automatically discovers cross-table temporal patterns.
Best for
Highest accuracy with minimal feature engineering. Captures publisher-user affinity, creative fatigue cycles, and campaign-level budget signals all at once.
Watch out for
Requires relational data in normalized tables. Not the right tool if you only have a single pre-aggregated feature CSV.
Key metric: RelBench benchmark: relational models score 76.71 vs 62.44 for single-table baselines on prediction tasks over multi-table data.
Why relational data changes the answer
CTR prediction is inherently a multi-entity problem. The click decision depends on the user (interests, device, context), the ad (creative type, campaign objective, frequency), and the publisher (content category, audience quality, placement). Flat-table models force you to collapse all of this into a single row of features per impression, destroying the relational structure that actually drives clicks. You end up with columns like 'user_avg_ctr_last_7d' and 'publisher_avg_ctr_electronics' -- static aggregates that miss the dynamic interplay between entities.
Relational models read these tables as a connected graph. They learn that User U003 on mobile has high affinity for Electronics ads specifically when shown on lifestyle publishers with premium inventory, and that this pattern strengthens on weekday mornings. These multi-hop, time-aware signals are exactly what manual feature engineering tries to approximate but rarely captures completely. On benchmarks like RelBench, relational approaches score 76.71 vs 62.44 for single-table baselines -- a gap that translates directly to millions in ad revenue at scale.
Predicting CTR from a flat feature table is like casting a movie by looking at each actor's headshot in isolation. You might pick individually talented people, but you will miss the chemistry between them. The relational graph is the screen test -- it shows you how the user, the ad creative, and the publisher context interact together, and that interaction is what makes the audience click.
How KumoRFM solves this
Graph-powered intelligence for advertising
Kumo builds a heterogeneous graph connecting users, ads, impressions, clicks, campaigns, and publishers. The GNN learns latent patterns like 'users who clicked similar creatives on related publishers' without manual feature engineering. PQL lets you express the prediction in two lines while Kumo automatically discovers the cross-table signals that traditional models require months of feature work to approximate.
From data to predictions
See the full pipeline in action
Connect your tables, write a PQL query, and get predictions with built-in explainability — all in minutes, not months.
Your data
The relational tables Kumo learns from
USERS
| user_id | segment | device_type | geo |
|---|---|---|---|
| U001 | Tech-savvy | Mobile | US-West |
| U002 | Bargain-hunter | Desktop | US-East |
| U003 | Luxury | Mobile | EU-West |
ADS
| ad_id | campaign_id | creative_type | category |
|---|---|---|---|
| A100 | CMP01 | Video | Electronics |
| A101 | CMP02 | Banner | Fashion |
| A102 | CMP01 | Native | Electronics |
IMPRESSIONS
| impression_id | user_id | ad_id | publisher_id | timestamp |
|---|---|---|---|---|
| IMP5001 | U001 | A100 | PUB01 | 2025-03-01 08:12 |
| IMP5002 | U002 | A101 | PUB02 | 2025-03-01 09:45 |
| IMP5003 | U003 | A102 | PUB03 | 2025-03-01 10:30 |
CLICKS
| click_id | impression_id | user_id | timestamp |
|---|---|---|---|
| CLK301 | IMP5001 | U001 | 2025-03-01 08:12 |
| CLK302 | IMP4990 | U002 | 2025-02-28 14:20 |
CAMPAIGNS
| campaign_id | advertiser | budget | objective |
|---|---|---|---|
| CMP01 | TechCorp | $500K | Conversions |
| CMP02 | FashionBrand | $200K | Awareness |
PUBLISHERS
| publisher_id | name | category | avg_ctr |
|---|---|---|---|
| PUB01 | TechNews | Technology | 2.1% |
| PUB02 | StyleMag | Fashion | 1.8% |
| PUB03 | LuxuryDigest | Lifestyle | 3.2% |
Write your PQL query
Describe what to predict in 2–3 lines — Kumo handles the rest
PREDICT BOOL(CLICKS.click_id, 0, 1, hours) FOR EACH IMPRESSIONS.impression_id
Prediction output
Every entity gets a score, updated continuously
| IMPRESSION_ID | USER_ID | AD_ID | CLICK_PROB |
|---|---|---|---|
| IMP5001 | U001 | A100 | 0.087 |
| IMP5002 | U002 | A101 | 0.023 |
| IMP5003 | U003 | A102 | 0.142 |
Understand why
Every prediction includes feature attributions — no black boxes
Impression IMP5003 -- User U003 x Ad A102
Predicted: 14.2% click probability
Top contributing features
User affinity for Electronics category
High
31% attribution
Publisher LuxuryDigest avg CTR
3.2%
24% attribution
Native creative on mobile
True
19% attribution
User clicked similar ads (last 7d)
4 clicks
15% attribution
Campaign frequency cap remaining
8 of 10
11% attribution
Feature attributions are computed automatically for every prediction. No separate tooling required. Learn more about Kumo explainability
PQL Documentation
Learn the Predictive Query Language — SQL-like syntax for defining any prediction task in 2–3 lines.
Python SDK
Integrate Kumo predictions into your pipelines. Train, evaluate, and deploy models programmatically.
Explainability Docs
Understand feature attributions, model evaluation metrics, and how to build trust with stakeholders.
Frequently asked questions
Common questions about ctr prediction
What is the best machine learning model for CTR prediction?
Graph neural networks that operate on relational data (users, ads, publishers, campaigns) outperform flat-table models by capturing cross-entity interactions. Traditional approaches like DeepFM and XGBoost require months of feature engineering to approximate what a GNN learns automatically from the table structure. The performance gap widens as your data becomes more relational.
How do you improve CTR prediction beyond feature engineering?
Stop collapsing your relational data into flat feature rows. Instead, let a graph model read your normalized tables directly. The biggest accuracy gains come from cross-entity signals -- publisher-audience affinity, creative fatigue propagation, campaign budget pacing -- that are nearly impossible to capture as hand-crafted features.
What data do you need for a CTR prediction model?
At minimum: impression logs, click events, and user profiles. For best results, add campaign metadata, publisher attributes, creative features, and device context. The power comes from joining these tables through their natural foreign keys, not from any single data source.
How does creative fatigue affect CTR prediction?
Creative fatigue is a cross-entity, temporal signal: the same ad shown to the same user segment on the same publisher decays in click rate over time. Flat models capture this crudely as 'days since creative launch.' Relational models track fatigue per user-creative-publisher combination, detecting when a creative still works for new audiences even as it fatigues returning ones.
What CTR lift can you expect from better prediction models?
Teams moving from hand-engineered feature tables to relational graph models typically see 5-15% CTR lift. For an ad platform serving billions of daily impressions, even a 5% improvement translates to $100M+ in annual revenue. The lift comes from capturing signals that flat models structurally cannot represent.
Bottom line: An ad platform serving 10B daily impressions that improves CTR prediction by 5% unlocks $120M in annual revenue. Kumo captures cross-entity signals between users, creatives, and publishers that flat feature tables miss entirely.
Related use cases
Explore more ad tech use cases
Topics covered
One Platform. One Model. Infinite Predictions.
KumoRFM
Relational Foundation Model
Turn structured relational data into predictions in seconds. KumoRFM delivers zero-shot predictions that rival months of traditional data science. No training, feature engineering, or infrastructure required. Just connect your data and start predicting.
For critical use cases, fine-tune KumoRFM on your data using the Kumo platform and Research Agent for 30%+ higher accuracy than traditional models.
Book a demo and get a free trial of the full platform: research agent, fine-tune capabilities, and forward-deployed engineer support.
