Pybandit: A Website Optimization Framework for E commerce SMBs
summary
Claims a practical way to run experiments with low traffic by shifting traffic to better pages as data arrives, reducing revenue lost to inferior variants. Implements an epsilon-greedy bandit in a pip-installable pybandit python library (v0.0.1) and demos a simulation with four arms where the bandit routes ~70–76% of traffic to the current winner versus 25% in fixed A/B allocation, raising conversion from ~4.0% to ~4.3% and estimating tens of thousands in extra revenue for sites with 5–10k monthly impressions. The demo supports hot add/remove of variants and live updating.
AI-written summary · about
chapters (10)
- 0:00Introduction to e-commerce SMB challenges
- 1:50Problem statement and current approaches
- 12:47Why Pi Bandit is needed
- 14:55Limitations of A/B testing
- 18:26How Pi Bandit works
- 21:21Advantages of Pi Bandit
- 24:12Demo and tool introduction
- 26:26Open-source Pi Bandit project
- 27:50Live demo simulation
- 37:11Conclusion and Q&A
related talks
21mBecoming the Multi-armed Bandit
36mTaking the Guesswork out of Performance Budgets
41mA/B Testing with Style
27mCustomer Driven Development : What, Why & How?
28mExperimenting your way to a better product
48mHow to use experimentation in the back-end
36mA/B Testing in Production
18mBuilding an AB testing analytics system with R and Shiny