Unbiased Learning from Biased User Feedback with Thorsten Joachims - TWiML Talk #207

Intervention Learning

Thorsten Joachims, a professor at Cornell University, explores the concept of learning to intervene in user behavior through machine learning and policy interventions. He explains that this approach is akin to medical treatments, where a specific action is taken, and the outcome is observed, similar to ad placements in online systems 1. Joachims emphasizes the importance of creating policies that cause desired behaviors, drawing parallels to controlled randomized trials in medicine.

What I really want is I want to make recommendations to the user, and I want to change behavior in a way that the user appreciates it.

--- Thorsten Joachims

This method shifts the focus from mere prediction to causal influence, aiming for interventions that users find beneficial 1.

   

Unbiased Learning

Joachims also delves into unbiased learning from biased user feedback, a key topic at the AI Summit. He discusses how biases are introduced in recommender systems and the role of inference techniques in mitigating these biases 2. By implementing appropriate logging policies, learning algorithms can become more robust against bias, enhancing their effectiveness.

This idea of learning from data is just fascinating.

--- Thorsten Joachims

Joachims shares his journey into machine learning, highlighting its potential to produce unforeseen results and advance AI capabilities 2.