Transcript Slides

Recommendation challenges at Amazon
Airstream use case
Houssam Nassif
AISTATS’15
Mission Statement
Airstream
Visual Browse Experience
Beautiful
• How to define and measure?
• Image parameters? Clicks? Hand curated?
Engaging
• How to engage users?
• Optimize for discovery or repeatability?
• What products to show?
Optimize for Users
Customers
• 244MM active accounts
• 14 countries
Recommendation at scale
• Personalized
• What metric to optimize?
• Milliseconds latency
Value (clicks, purchases)
Temporal Variations
Time (in hours)
Customers
Zero-Inflated Distribution
Engagement
How to Recommend?
Multi-armed
bandits
Collaborative
filtering
Regression
analysis
Deep
learning
…
Need for Diversity
How to Diversify?
Challenges
• How to learn ideal mix?
• How to balance between diversity and
metric of interest?
Possible solutions
• Determinantal Point Processes
• Submodular Functions
ML Teams @ Amazon
S9
ML Seattle
Toronto
Dev Center
ML Berlin
ML Bay Area
A2Z
A9
ML Bangalore