Abstract:
|
The group fused lasso is a penalized linear regression method for data such as multivariate time series, gene sequencing data, and multichannel images, that display both group- and graph-structure. Applications of this powerful method include segmentation, change-point detection, prediction, and signal recovery. This talk explores the computation of sparse group fused lasso (SGFL) with high-dimensional time series. Although SGFL can be implemented with standard techniques for nonsmooth convex optimization, these techniques do not scale to high dimension and may also suffer from lack of interpretability. To address these limitations, I introduce a fast SGFL algorithm that adaptively performs local and global optimizations using block coordinate descent, majorization/minimization, and steepest subgradient descent. In addition to its speed, the proposed approach strictly enforces structural constraints, e.g., global change points, and guarantees convergence to a global solution. The new method is compared to the state of the art in a numerical study and illustrated with resting-state fMRI data.
|