Linear Regression with Unknown Truncation Beyond Gaussian Features
A breakthrough algorithm solves a long-standing limitation in truncated regression models.
Deep Dive
Researchers have developed the first polynomial-time algorithm for truncated linear regression with an unknown survival set, a problem studied since 1897. Previous methods required strong assumptions like Gaussian features and had exponential runtimes. The new algorithm works with sub-Gaussian features and runs in poly(d/ε) time, breaking a major computational bottleneck. This was achieved via a novel subroutine for learning unions of intervals from positive-only examples under smoothness conditions.
Why It Matters
This enables more accurate and efficient modeling of real-world data where observations are systematically missing or censored.