【Job talk】Statistical analysis of greedy algorithms: Unit-root time series and scalable multi-task learning

  • 研討會日期 : 2024-01-11
  • 時間 : 10:45
  • 主講人 : Mr. Shuo-Chieh Huang (黃碩傑)
  • 地點 : Conference Room B110
  • 主持人 : Professor Chu-An Liu
  • 演講者簡介 : Mr. Huang is expected to receive his Ph.D. in Economics from the University of Chicago in 2024. His research fields are Distributed learning, Machine learning, Big dependent data, Optimal transport, Econometrics. He is applying for a position of the Institute of Economics, Academia Sinica now.
  • 演講摘要 : In this talk, I will demonstrate the usefulness of greedy algorithms both in highly persistent time series and in big, distributed computing architecture. First, we propose a greedy-based algorithm, FHTD, for consistent variable selection of the high-dimensional unit-root ARX model, in which a fully general but unknown unit-root structure is allowed. Second, for estimating a multi-task linear regression with feature-distributed (or vertically partitioned) data, we employ the two-stage relaxed greedy algorithm (TSRGA). Because of its low communication complexity, which does not scale with the ambient dimension, TSRGA is computationally attractive in this setup. In both cases, the key theoretical ingredient is to characterize the rate of convergence along the iteration path. When applied to real-world economic data, the methods outperform commonly-used benchmarks.