The evaluation metric partitions the dataset into 3336 subsets (278 neighborhoods times 12 months), and for each subset, compares the distribution over 174 incident types with the ground truth. This effectively asks for a full contingency table. When the privacy parameter epsilon is sufficiently large, using the Laplacian mechanism to obtain a full contingency table suffices. However, when epsilon is small, the full contingency table is too noisy, and the given public data can provide a better estimate. Our approach privately chooses which method to use.
Prize amount: $7,000
Team members: Ninghui Li, Zitao Li, Tianho Wang