These are slides presented at MLconf in San Francisco, November 14, 2014. I share the approach to real-time machine learning for recommender systems developed at if(we). We achieve rapid iterative cycles by adhering to a strict approach to structuring and accessing our data, as well as to building the online features that comprise our models. These developments support teams of data scientist and data engineers, who work together to solve complex recommendation problems. We also introduce the Antelope Realtime Events framework, an open source demonstration application which derives from our scalable proprietary software stack.

1. Agile Machine Learning for Real-time
Recommender Systems
Johann Schleier-Smith
CTO, if(we)
@jssmith johann@ifwe.co github.com/ifwe

2. what it should look like

3. 1. Gain understanding of machine learning
2. Gain understanding of the product usage
3. See opportunity to make the product better
4. Create training data
5. Train predictive models
6. Put models in production
7. See improvements

4. what it often looks like

5. 1. Gain understanding of machine learning
2. Gain understanding of the product usage
3. See opportunity to make the product better
4. Pull records from database to create interesting
features (usually aggregates)
5. Train predictive models
6. Go implement models for production
7. See improvements

6. 1. Gain understanding of machine learning
2. Gain understanding of the product usage
3. See opportunity to make the product better
4. Pull records from database to create interesting
features (usually aggregates)
5. Train predictive models
6. Go implement models for production
7. See improvements
3-6
months

7. 1. Gain understanding of machine learning
2. Gain understanding of the product usage
3. See opportunity to make the product better
4. Pull records from database to create interesting
features (usually aggregates)
5. Train predictive models
6. Go implement models for production
7. See improvements Cool!
Wa s i t w o r t h i t ?

8. • Profitable startup actively pursuing big
opportunities in social apps
• Millions of users of existing brands
• Thousands of social contacts per second

9. real-time
recommendations
challenges

10. Tagged dating feature
• >10 million candidates
to select from
• >1000 updates/sec
• Must be responsive to
current activity
• Users expect instant
query results

11. implementation
pain points

12. • Data scientist hands model description to
software engineer
• May need to translate features from SQL to Java
• Aggregate features require batch processing
• May need to adjust features and model to
achieve real-time updates
• Fast scoring requires high-performance in-memory
data structures

13. time for
new thinking

14. one way that
works better

15. !
!
!
4. Pull records from database to create interesting
features (usually aggregates)
5. Train predictive models
6. Go implement models for production

16. Create interesting features
Train predictive models
Put models in production

17. Create interesting features
Train predictive models
Put models in production

18. one right way to data
event history

19. History.
filterTime(start, PLUS_INFINITY).
foreach {
e: Event => model.update(e)
}

20. everything is an event

21. Bob registers
Alice registers
Alice updates profile
Bob opens app
Bob sees Alice in recommendations
Bob swipes yes on Alice
Alice receives push notification
Alice sees Bob swiped yes
Alice swipes yes
Alice sends message to Bob

22. writing the model

23. class MyModel {
def update(e: Event) { … }
def topN(ctx: Context, n: Int) = { … }
}

24. models are all
about features

25. class MyFeature {
def update(e: Event) { … }
def score(ctx: Context,
candidateId: Long): Double = { … }
}

26. model training

27. History.
filterTime(start, PLUS_INFINITY).
foreach {
e: Event => {
writeTrainingData(outcome(e),
model.features(context(e))
model.update(e)
}
}

28. live demo
Kaggle competition
with Best Buy data
https://www.kaggle.com/c/acm-sf-chapter-hackathon-small

29. product update events
{
“timestamp” : “2012-05-03 6:43:15”,
“eventType” : “ProductUpdate”,
“eventProperties” : {
“sku” : “1032361”,
“regularPrice” : “19.99”,
“name” : “Need for Speed: Hot Pursuit”,
“description” : “Fasten your seatbelt and
get ready to drive like your life depends
on it...”
...
}
}

30. product view events
{
“timestamp” : “2011-10-31 09:48:46”,
“eventType” : “ProductView”,
“eventProperties” : {
“skuSelected” : “2670133”,
“query” : “Modern warfare”
}
}

31. demo
Try it yourself, code and instructions at:
https://github.com/ifweco/antelope/blob/master/doc/demo.md

35. 1. Gain understanding of machine learning
2. Gain understanding of the product usage
3. See opportunity to make the product better
4. Create training data
5. Train predictive models
6. Put models in production
7. See improvements
Fast cycles!!

37. • All data in form of events – no exceptions!
• Roll through history to generate training examples
• Sample training data carefully to avoid feedback
• Model is static while features are live and personal
• Use interesting features with boring algorithms
• Expressiveness > performance > scalability
github.com/ifwe/antelope
@jssmith