This document discusses using big data and machine learning techniques on AWS for content recommendations. It describes three common approaches: search with boosting which adjusts search rankings based on popularity signals; collaborative filtering which identifies similar users and items; and neural networks which use historical user events to create a model that predicts favorites. It also introduces Amazon DSSTNE (Deep Scalable Sparse Tensor Network Engine) for automating GPU-accelerated training and prediction at scale for recommendation systems.

1. Big Data, Analytics, and
Content Recommendations on AWS
Mike Limcaco| Senior Manager – Solutions Architecture
Amazon Web Services

2. http://mashable.com

3. Search
Watch
Listen
Play
Download
Purchase
Rate It
Review It
Sharing
Tagging
Bookmarking

4. Content
AudienceOperations

5. What are the common asks?
Content
•Top Content
•Engagement
• Plays per session
• Drop-off
•Referral path
•Recommendations
Audience
• Acquisitions
• Churn
• Where, when,
who
• Segmentation
• Cohorts
Operations
• How much
buffering
• Best CDN
paths
• Top devices
• Uniques per
platform
Other
•Monetization
•Ad Spend
•Social Media
• Mentions
• Sentiment Analysis
•A/B Feature Testing
•Talent Management

6. (Some) AWS Big Data Services
Amazon S3
Internet scale
storage
Amazon Elasticsearch
Hosted Elasticsearch
Distributed Search
Engine
Amazon EMR
Hosted Hadoop
compute framework
• Data transformation
• Aggregations
• Predictions
• Discovery
• Visualization
• Content Lake
• Raw signals
• Clickstream
• Profiles & models

7. Show me …

8. Content
Recommendations

10. Incrementally … magical*
1. Search with Boosting
2. Collaborative Filtering
3. Neural Networks
A
B
C
precision means greater engagement*

11. Search With BoostingA
1. Capture Audience Signal Data
2. Record aggregate “Popularity” in the search catalog
 Counts of Views | Downloads
 Trending Sentiment (Positive | Negative)
3. Query Search Engine
4. Use content metadata + “Popularity” to refine
search ranking
5. Enjoy recommendations!

12. or

13. vote = count(Event: title, device, time)
VOD Catalog
title: “Toy Dogs and Their Owners”
date: 2014
content: “Cute but surreal look at … ”
votes: 25
title: “The Usual Suspects”
date: 1995
content: “A sole survivor …”
votes: 85
title: “Star Wars 17”
date: 2026
content: “Yoda force ghost returns … surreal”
votes: 99

14. GET /catalog/movies/_search
Search engine adjusts
rankings based on additional contribution of dynamic field “votes”
VOD Catalog
title: “Toy Dogs and Their Owners”
date: 2014
content: “Cute but surreal …”
votes: 25
title: “The Usual Suspects”
date: 1995
content: “A sole survivor …”
votes: 85
title: “Star Wars 17”
date: 2026
content: “Yoda force ghost returns … surreal”
votes: 99
Rank #1

15. Collaborative Filtering
1. Capture Audience Signal Data
Create history of user and item preferences
2. Estimate similar users and items
3. Record these in Search Engine
4. Query Search Engine with User History
5. Enjoy recommendations!
http://www.slideshare.net/tdunning/recommendation-techn
B

16. users
Media
platforms
Mobile
Search
Play
Buy
Rate
Recommendations

17. mike,view,movie-a
mike,view,movie-b
mike,view,movie-c
mike,buy,movie-b
chris,view,movie-b
chris,buy,movie-d
…
movie-b movie-c:2.772588722239781
movie-a:2.772588722239781
movie-d ….
Indicators
(“Items Similar To This….”)
% mahout spark-itemsimilarity
-i input-folder/data.txt
-o output-folder/
--filter1 buy -fc 1 -ic 2
--filter2 view

18. Step 1: Logs  History Matrix
User1 Thing1
User2 Thing2
User3 Thing3
User2 Thing4
User5 Thing1
User1 Thing2
User1 Thing3
Mike
Jon
Mary
Phil
Kris
Logs History Matrix

19. Step 2: Estimate Similar Things
History Matrix
2 8
2 4
8
4
Item-Item Matrix

20. Step 3: Reduce to Interesting Pairs
2 8
2 4
8
4
Item-Item Matrix
LLR
Indicators
(“Items Similar To This….”)

21. Step 3: Reduce to Interesting Pairs
Indicators
(“Items Similar To This….”)
Items Similar To This

22. Step 4: Store Indicators in a Search Engine (BATCH)
Superman Highlander,
Dune
Star Wars Raiders,
Minority
Report
Highlander Superman
Mulan Home Alone,
Mermaid
Star Trek …
… …
4587 223, 5234
748 5345, 235
12 8234
245 9543, 7673
3456 4587
… …
Index

23. Indicators

24. Step 5: Query Search Engine w/ User History (REALTIME)
748 Star Wars 45, 235
12 Highlander 8234
245 Mulan 9543, 7673
4587 Superman 12, 5234
3456 Star Trek 2458 …
Query
“12”
5345
3456
12

25. Neural NetworksC
1. Capture Audience Signal Data
Create history of user and item preferences
2. Create a model (GPU) which captures the
relationships between users and items
3. Use the model to score (GPU) users and predict
their favorite items
4. Enjoy recommendations!

26. https://lazyprogrammer.me

27. Historical
User Events
(Watch,
Buy,
Subscribe)
Titles / Categories
Predictions
(Watch,
Buy,
Subscribe)

28. Historical
User Events
(Watch,
Buy,
Subscribe)
Titles / Categories
Predictions
(Watch,
Buy,
Subscribe)
Iterating and finding the right mix of weights (influence) that describes observed patterns in aggregate

30. Amazon DSSTNE
Deep Scalable Sparse Tensor Network
Engine
https://github.com/amznlabs/amazon-dsstne

31. Why DSSTNE
• Automated management of multi-GPU parallelism
• Scale out training
• Scale out predictions
• Optimised for Sparse Data Efficiency
• Lots of users and lots of products … but relatively small
overlap
https://github.com/amznlabs/amazon-dsstne

32. % train -c config.json -i gl_input.nc -n the-computed-model.nc …
(…. A little later …)
% predict … -k 10 -n the-computed-model.nc …
-r input-observations.txt
-s output-recommendations.txt

33. Sample Predictions (MovieLens 20M)
Training Time on g2 ~ 1:45s
Scoring 130K users ~ 20s
User 22 has great taste!

34. CG1 G2 P2
Processor X5570
(Nehalem)
E5-2670
(Sandy Bridge)
E5-2686 v4 (Broadwell)
GPU 2x Nvidia Tesla Fermi
M2050
4x Nvidia GRID 8 Nvidia K80 (2 GK210
GPU’s each)
CUDA Cores 896 6144 39,936 (23 TFlops)
GPU RAM (GB) 6 16 12
Cores 8 16 32
Clock Speed 2.9 – 3.33 2.6 – 3.33 2.7
RAM (GB) 22.5 60 732
RAM/pCore 2.81 3.75 22.8
Storage 1.68TB Magnetic 240GB SSD EBS Only
Networking 10Gbit 10Gbit 20Gbit
Cluster GPU

35. Phase 1:
Training on GPU
Phase 2 ...N:
Scoring (Predicting) on GPU

36. https://aws.amazon.com/blogs/big-
data/generating-recommendations-at-amazon-
scale-with-apache-spark-and-amazon-dsstne/

37. Thank You!