Zoom webinar link

Data valuation frameworks like data Shapley values quantify how individual data points contribute to predictive performance in machine learning. While useful for fair valuation, organizations could use them to assign zero value to some individuals, reducing their incentive to consent to data sharing. To address this, we propose an Information Disclosure Game between a Data Owner (DO) and Data Consumer (DC). Rather than sharing raw data, the DO incrementally reveals information with Laplacian noise under differential privacy. Focusing on $k$-nearest neighbors, we simulate DC strategies using data Shapley values and multi-armed bandit exploration. Experiments on a Yelp review task suggest that data valuation imposes real costs on the DC’s selection process.

(with Gilles Caporossi and Sébastien Gambs)