zaid-g / ccb_tutorial

Contextual multi-armed bandit recommender system using Vowpal Wabbit

• 👋 Hi, I’m Zaid (pronounced Zayd)
• 🌱 My main interests are in machine learning (especially RL, LLM's, multi-armed bandits) and economics
• 👀 I’m looking to collaborate on impactful projects matching one or more of my interests
• 📫 zaidg at uw dot edu

ccb_tutorial's Issues

CCB Complete Cycle

Hi Zaid, I have gone through your blog on Conditional contextual bandits. It was a great blog for someone who has just started using the VW. I wanted to understand the entire flow of CCB when we deploy it in production. How do we re-train the model and how to evaluate the performance of the re-trained model.
So lets say we warm start the CCB model using the logged data ( fixed position of actions ) and deploy that model in production for a small traffic of user. Now we will receive rewards based on the predictions my model made. We collect the data lets say for couple of hours and using that data we again trained our model. and this cycle will keep on going. Is. my understanding correct for the overall flow. How do I evaluate the performance of my model in this case?
Thanks