Experience Optimization
using
Contextual Bandits
Trey Causey
zulily
@treycausey
s/clos/experiment/g
Moe: "You could flash-fry a
buffalo in 40 seconds."
Homer: "Awwww, but I want it now!"
Experiments
- High-power experiments may take too long
- Can't adjust to rapidly changing content
- Experiment's lifetime == content's lifetime
- May not want to select one treatment
Multi-armed bandits
- Reinforcement learning
- Exploration vs. exploitation
Contextual bandits
- Incorporate features of user and content
- The best experience for you, right now
- ... even if that changes.
Development
- Developed in Python
- Tested with Monte Carlo simulation
Implementation
- Capacity
Has to handle thousands of trials/second
Has to handle thousands of trials/second
- Latency
Called on page load to determine experience
Called on page load to determine experience
- Implementation
In production in Java