ADM+S researchers finalists in UC Berkeley Prosocial Ranking Challenge
Author Kathy Nickels
Date 17 May 2024
A team led by ADM+S researchers Dr Aaron Snoswell, Distinguished Professor Jean Burgess and William He from the new GenAI lab at QUT in collaboration with ADM+S Associate Investigator Dr Daminao Spina (RMIT University) and Dr Tariq Choucair from QUT have been announced as one of nine finalists in the Prosocial Ranking Challenge.
The Prosocial Ranking Challenge, hosted by the Center for Human-Compatible AI at UC Berkeley, awards $60,000 in prizes to build better social media algorithms.
The challenge tests ways to mitigate problems or harms associated with social media algorithms, it also seeks to demonstrate new ways to design systems toward socially desirable ends.
The team’s approach is based on the concept of Search Result Diversification.
Dr Snoswell said, “We aim to mitigate political and other forms of polarization by exposing users to diverse opinions and content.
“To do this, we use a Large Language Model (LLM) to simulate personas with diverse political perspectives, rank the users’ social media feed according to each of these personas, and then combine these partisan rankings with an award-winning fairness-preserving algorithm that balances the range of opinions present in the news feed.”
As finalists, the team has one month to create a production-ready version of their algorithm that meets performance and security requirements. If successful the algorithm will be selected to take part in a large-scale trial to evaluate the real-world performance of the algorithms, with results published in 2025.