Recommendations at SEA with Olivier Jeunen and Ben Carterette

***
IMPORTANT: You will be able to view the Zoom link once you 'attend' the meetup on this page.
***

We're back from our Summer Break. This month's SEA features Olivier Jeunen (U. Antwerp) and Ben Carterette (Spotify).

17.00: Olivier Jeunen (U. Antwerp)
Title: Advances in Bandit Learning for Recommendation: Pessimistic Reward Models & Equity of Exposure
Abstract: The “bandit learning” paradigm is an attractive choice to recommendation practitioners, because it allows us to optimise a model directly for the outcomes driven by our recommendations. Nevertheless, because we do not observe the outcomes of actions the system did not take, learning in such a setting is not straightforward. In the off-policy setting, where we learn from a fixed dataset of logged recommendations and their outcomes, selection bias of this form is prone to lead to problems of severe over-estimation. The first part of this talk will introduce Pessimistic Reward Models, and show how they can alleviate this bias, leading to significantly improved recommendation performance. In the on-policy setting, we learn from the outcomes of our own actions, and several provably optimal algorithms exist. But is blindly optimizing a model for some user-focused notion of reward always what we want? The second part of this talk will discuss the “Equity of Exposure” principle in relation to Top-K recommendation problems, introducing an Exposure-Aware Arm Selection algorithm that can significantly improve fairness of exposure with a minimal impact on reward.

17.30: Ben Carterette (Spotify)
Title: Measuring Listener Delight
Abstract: At Spotify, listeners' delight with their recommendations and audio sessions is one of our major considerations. Delight---as distinguished from satisfaction or engagement---is not easy to measure, even in a user study, and much less so when limited to the in-app implicit and explicit feedback signals that our listeners provide. One reason for the difficulty is simply the challenge of interpreting those signals intelligently. What does it mean when a listener skips a track, for example? How should we interpret repeated skips in a long session? But even when we understand this feedback well, leveraging it for offline evaluation and training only raises further challenges. Taking listener delight as our north star, I will talk about some of our recent work towards understanding user feedback, using it for improved training and evaluation offline, and the barriers that still remain before we can truly say we have a robust scientific understanding of how our recommendations bring joy.

---
This is (still) an online event. The URL will be shared close to the day of the event. All times are Amsterdam times.