Recommendations systems are a ubiquitous part of everyday life helping us navigate the immense amounts of information available online. But following recommendations can lead to unintended consequences such as polarization and radicalization.
We propose a way of evaluating recommendations system by looking at the degree of agency they grant to individuals using the system. People have agency within a system if they can modify how the algorithm perceives their preferences. In our research we propose a way of quantifying the degree of agency in recommenders and argue that it is a promising new directions for evaluating algorithms that interact with humans.
For more detail about this research check out our short paper: https://people.eecs.berkeley.edu/~sarahdean/stochastic_reachability.pdf
To see more recent work in Participatory Approaches to Machine Learning check out the workshop at ICML 2020: https://participatoryml.github.io/
To see more of our other work see our research websites:
Sarah Dean: https://people.eecs.berkeley.edu/~sarahdean/
Mihaela Curmei: https://mcurmei627.github.io/research/