Exploration: Measurements and Systems

Dr. Minmin Chen (Google DeepMind)


Human curiosity compels us to explore and understand the unknown. The information retrieval and recommendation systems we relied upon daily for information acquisition and decision making however often fall prey to the "closed feedback loop". They reinforce familiar and popular choices and leave behind a large base of worthy content and creators, limiting our discovery.

In this talk, I will dive deep into exploration as a key strategy to combat this bias and unlock the potential of fresh and under-explored content in recommender systems. I will discuss new experiment frameworks we designed to rigorously quantify the values of exploration on corpus, creators, and the users. I will share our journey in building dedicated exploration stacks within a large-scale industrial recommendation platform to incubate and provide the initial exposure to fresh and tail content, shielding them from the fierce competition from the main recommendation flow, as well as integrating new exploration algorithms into the core recommendation flow themselves, lowering the growth barriers for these hidden gems.

Today, we are in another revolution where LLMs are poised to change how we build every system, but critical challenges remain. I will also cover a recent proposal on a hybrid approach to leverage LLMs and classic recommendation models together to address the challenging tasks of helping users explore the unknowns. 

Speaker Biography:

Minmin Chen is a principal research scientist at Google DeepMind, leading efforts in building conversational AI systems through personalization and RL.  She received her PhD from Washington University in St. Louis. Her main research interests are in reinforcement learning and bandit algorithms and their applications to recommendation and assistive systems. She received 2024 WSDM best paper award for her work in exploration in recommender systems. She is a guest editor of Machine Learning Journal and Area Chair for Neurips, ICML, ICLR and ACM RecSys.