Learning and Planning Towards AI for Social Good

 AI for Social Good (AI4SG) is a research theme that uses and advances AI to im?prove the well-being of society. We introduce three lines of work that center around learning and planning to address real-world challenges in cybersecurity, food waste and security, and environmental conservation. For cybersecurity, we provide a learning and planning pipeline for generic cyber deception and an algorithm to counter watering-hole attacks. In food waste and food security, we develop a pre?dictive model for the rescue claim status and an online learning and planning algo?rithm for volunteer engagement through push notifications. We also ran a random?ized controlled trial for our algorithm to show significant improvement in the real world. For environmental conservation, we develop a natural language processing?based media content monitoring system to provide early warning of infrastructure projects that might pose harm to conservation efforts. The system leverages ac?tive learning and learning with noisy labels algorithms to address challenges in applied learning and planning applications. The tool has been deployed in multi?ple places around the world, monitoring over 60,000 conservation sites worldwide since February 2022. Distilling lessons learned from these projects, we propose ban?dit data-driven optimization, the first iterative learning and planning framework to rigorously address the pain points in practical prediction-prescription workflows in lots of social good projects across application domains