Synthetic Data Generation for CTR Optimization

This document explains how to bootstrap the CTR optimization system with realistic synthetic data when historical ad serving data is not available.

Overview

The CTR prediction model requires 7 days of historical data. In development environments without real traffic, synthetic data generation provides realistic ad serving events with embedded CTR patterns.

Embedded Patterns

Pattern Type	Values	Impact
Device CTR	Mobile: 2.5%, Desktop: 1.2%, Tablet: 1.8%	Baseline performance
Time of Day	Evening (6-10 PM): +60-70%, Business (9-5 PM): +20-30%	Engagement multiplier
Day of Week	Wednesday: Peak, Weekdays > Weekends, Sunday: Lowest	Traffic distribution
Line Items	CPC campaigns: +20%, CPC + Mobile: +30%	Performance boost

Usage

Automated Bootstrap (Recommended)

# Start the CTR predictor service
docker compose up ctr-predictor -d

# Run the bootstrap script
./scripts/bootstrap-ctr-optimization.sh

This script generates 7 days of data, trains the model, and verifies predictions.

Manual Steps

# 1. Generate synthetic data
curl -X POST http://localhost:8000/generate-synthetic-data?days=7&impressions_per_day=10000

# 2. Train the model
curl -X POST http://localhost:8000/train \
  -H "Content-Type: application/json" \
  -d '{"days_back": 7, "min_impressions": 100}'

# 3. Test prediction
curl -X POST http://localhost:8000/predict \
  -H "Content-Type: application/json" \
  -d '{
    "line_item_id": 2,
    "device_type": "mobile",
    "country": "US", 
    "hour_of_day": 20,
    "day_of_week": 2
  }'

Verify Data Generation

-- Check total events generated
SELECT 
    event_type,
    count(*) as events,
    toDate(min(timestamp)) as first_date,
    toDate(max(timestamp)) as last_date
FROM events 
GROUP BY event_type;

-- CTR by line item
SELECT 
    line_item_id,
    countIf(event_type = 'impression') as impressions,
    countIf(event_type = 'click') as clicks,
    clicks / impressions as ctr
FROM events 
GROUP BY line_item_id
ORDER BY line_item_id;

Data Characteristics

Metric	Value	Description
Duration	7 days	Historical data period
Volume	~10,000/day	Impressions (configurable)
Overall CTR	~2.5%	Varies by context
Device Split	60% mobile, 30% desktop, 10% tablet	Traffic distribution
Peak CTR	~4.5%	Mobile CPC evening
Baseline CTR	~1.0%	Desktop CPM morning
Noise	±20%	Random variation

Expected Model Performance

Context	Boost	Description
Mobile + CPC + Evening	1.5x	High CTR context
Desktop + CPM + Morning	1.0x	Baseline performance
Tablet + CPC + Business	1.3x	Good performance

Testing CTR Optimization

Once bootstrapped, test the full optimization pipeline:

# 1. Enable CTR optimization
export CTR_OPTIMIZATION_ENABLED=true

# 2. Send ad requests with different contexts
curl -X POST http://localhost:8787/ad \
  -H "Content-Type: application/json" \
  -d '{
    "device": {"devicetype": 1},  // Mobile
    "user": {"id": "test-user"},
    "imp": [{"id": "1", "tagid": "test-placement"}]
  }'

Monitoring

Check logs for CTR optimization:

docker compose logs openadserve | grep "CTR optimization applied"

Example: INFO CTR optimization applied line_item_id=2 base_ecpm=3.0 boost_multiplier=1.5 optimized_ecpm=4.5

Updating Synthetic Data

To regenerate with different patterns:

1. Modify synthetic_data.py parameters
2. Rebuild CTR predictor: docker compose build ctr-predictor
3. Re-run bootstrap: ./scripts/bootstrap-ctr-optimization.sh

Production Transition

As real traffic accumulates:

1. Monitor real vs synthetic data quality
2. Retrain model daily: curl -X POST http://localhost:8000/train
3. Gradually increase min_impressions threshold
4. Disable synthetic generation once sufficient real data exists

Troubleshooting

Issue	Solutions
No events generated	Check ClickHouse connectivity, verify events table exists, check logs
Model training fails	Ensure data generated first, verify >1000 events, check permissions
Predictions not working	Confirm model loaded (/model-info), verify CTR optimization enabled