This academic paper focuses on differentially private submodular maximization, a technique crucial for data summarization in scenarios involving sensitive information. It explores algorithms that can maximize submodular functions—which capture diminishing returns, useful for tasks like feature selection and data summarization—while also adhering to differential privacy, ensuring individual data points cannot significantly influence the outcome. The research presents novel privacy-preserving greedy algorithms and analyzes their performance under various constraints (cardinality, matroid, p-extendible systems), demonstrating that these approaches can achieve competitive accuracy compared to non-private methods while safeguarding privacy. Experiments on location data and health features validate the practical utility of these techniques.
