Deploying Predictive Models

Full post here:

Transcript

1. SF Data Mining January 26, 2015 Deploying predictive models Nick  Elprin   Domino  Data  Lab   dominodatalab.com
2. Who am I? SF Data Mining January 26, 2015 • Founder of Domino Data Lab, a software platform for enterprise data science
 
 
 • Previously built analytical software at a big hedge fund
 
 
 • BA, MS in computer science
3. Motivation SF Data Mining January 26, 2015 Build predictive models Build production software systems Different languages good for different tasks
4. Motivation SF Data Mining January 26, 2015 Organizational design friction Model improvements Data Scientists Software Engineering Delayed because of: • Integration / porting of logic • Out-of-phase release cycles • Mismatched priorities
5. Solution SF Data Mining January 26, 2015 Publish Consume Data scientists create predictive models and publish  them  to  Domino. Domino provides a secure, low- latency infrastructure for hosting predictive models as  web  services Developers can invoke models from general purpose languages by making simple  HTTP  calls • Failover • Security • Logging • Seamless updates • etc
6. Demo SF Data Mining January 26, 2015
7. Production concerns SF Data Mining January 26, 2015 •Very low latency •Zero-downtime upgrades •High availability •Reproducibility •Logging •Security
8. Best practices SF Data Mining January 26, 2015 •Separate training, initialization, and prediction •Make your prediction functions thread-safe •Don’t mutate any shared state •Leverage persistence/serialization tools (e.g., pickle)
9. Use cases SF Data Mining January 26, 2015 •Lease / loan approval •Recommendation systems •Music, books, products, cars, etc •Insurance •Quoting premiums; claims estimates
10. SF Data Mining January 26, 2015 dominodatalab.com blog.dominodatalab.com @dominodatalab Check us out Webinar on parallel 
 programming in R and Python. Jan 28, 10:30am dominodatalab.com/webinar