Upload
datadays
View
87
Download
1
Embed Size (px)
Citation preview
CLICK PREDICTIONS FOR 500M MOBILE DEVICES
© 2008-2014
ADVERTISING IS A DATA BUSINESS
Refinement of a commodity: pixel real-estate
Value creation: delivering advertiser message to right client in context
KPI: quality and price
Telekom Shop Bonn-CityFriedensplatz 453111 Bonn
0.5km
© 2008-2014
ADVERTISING IS A BIG DATA BUSINESS
>1B unique devices
~40B impressions per day traded through RTB WW
40TB of new data per day
~500k impressions per second
Latency: 100ms
© 2008-2014
GEOGRAPHIC DISTRIBUTION
50msRTT
50msRTT50ms
RTT50msRTT
© 2008-2014
TECH CHALLENGE: SCALE AND COST
Prediction latency: 3ms
Volume: 500k impressions per second
= 300 predictions per core
= 1,500 cores = 150-200 servers for bidding
Cloud Cost:$200,000 per month
© 2008-2014
PREDICTION
Multiple concurrent predictive models
Likelihood to act (click, land, browse, buy)
Demographic groups
Sociographic groups
Interest groups
Fraud detection
Site classification
100’s of TB input data sets for training
Runs for many hours a day on even more machines
Some refreshed hourly
© 2008-2014
DATA ENRICHMENT
Multiple enrichments per bid request
User
Location
Device
Weather
10’s of Billion records with <1ms access latency
Data store with >1M IOPs/second
© 2008-2014
PREDICTIVE MODEL
Predictive model has substantially cost impactEvery impression individually priced based on KPI estimation (e.g., click)
Basic filtering is insufficient
Continuous model update in multiple time scales
1. Model refinement in real time
2. Full retraining at longer time scales (hour, day, week)
© 2008-2014
REAL-TIME BIG-DATA BUSINESS
Technology and process
Blue print invented by financial world
Layer-cutting integration
Human touch points with distributed ownership
Scales when done right
Real-Time Decisioning
Data Acquisition
Storage
Batch Analytics
Model Testing
Time Series and Cubes
Dashboard and
Reports