Abstract: Accurate and reliable prediction of outcomes is essential for decision-making across various domains such as healthcare, finance, and autonomous driving. Today’s data-rich landscape offers unprecedented opportunities for enhancing predictions to inform actionable decisions. While various statistical and machine learning (ML) approaches have been improving data-driven predictions and decisions, challenges remain in providing both a robust theoretical foundation and versatility to handle large, complex data. To address this critical gap, my research aims to advance the decision-making process by integrating ML algorithms with statistical principles. This integration will not only strengthen the fundamental aspects of ML approaches, but also create versatile, reliable ML methods, ultimately empowering decision-makers in real-world scenarios.

In this seminar, I will present two research projects demonstrating the synergy between statistics and ML. The first focuses on developing predictive methods resilient to data corruption and complex response variables, by addressing errors-in-variables regression with a metric-space-valued response. This work highlights the synergy between ML approach and statistical modeling for versatile prediction with error mitigation. The second focuses on increasing decision reliability through calibrated predictions. Reexamining classical concepts in calibration, we propose a holistic viewpoint for probability calibration, yielding a distribution-free probability calibration method with judicious parameter selection. These studies underscore the potentials of integrating statistical rigor and advanced ML algorithms to elevate the accuracy, reliability, and efficiency of decision-making processes in diverse applications. Time permitting, I will conclude the talk by sharing a glimpse into my ongoing efforts and future research directions.