Interpretable Learning at Scale: Getting Optimal Sparse Models with GPUs
As machine learning models continue to grow in size and complexity, they often become harder to understand, troubleshoot, and constrain using domain knowledge. In high-stakes settings such as healthcare and scientific discovery, practitioners frequently prefer models that are both accurate and interpretable. Yet learning such models can be computationally challenging, since enforcing sparsity often leads to mixed-integer optimization problems. In this talk, I will discuss recent progress on solving a class of such problems–sparse generalized linear models–to global optimality on modern large-scale datasets. Traditional mixed-integer optimization methods often struggle on these problems and make limited use of GPU hardware. To address this gap, we reformulate the perspective relaxations as composite optimization problems and develop a unified proximal framework that is GPU-friendly and linearly convergent. The resulting methods exploit problem structure across multiple scales to dramatically accelerate dual-bound computation. On large-scale datasets, we obtain orders-of-magnitude speedups in dual-bound computation and substantially improved scalability over state-of-the-art solvers such as Gurobi and MOSEK. More broadly, this work suggests a new path for combining interpretability, discrete optimization, and modern hardware for data-driven decision-making. This is joint work with Andrea Lodi and Soroosh Shafiee.
Bio: Jiachang Liu is an assistant research professor at the Center for Data Science for Enterprise and Society. His research interests include (1) creating interpretable and trustworthy ML solutions for high stakes decisions making, in domains such as healthcare, criminal justice, and finance; (2) designing efficient discrete and continuous optimization techniques to solve related optimization problems, which are usually nonconvex and have a combinatorial nature; and (3) building open-source and user-friendly software packages for the broad data science community. The long-term goal is to let humans and machines seamlessly collaborate and complement each other. Prior to joining Cornell, Liu completed his Ph.D. in electrical and computer engineering at Duke University.