A Generalized Adaptive Joint Learning Framework for High-Dimensional Time-Varying Models

In modern biomedical and econometric studies, longitudinal processes are often characterized by complex time-varying associations and abrupt regime shifts that are shared across correlated outcomes. Standard functional data analysis (FDA) methods, which prioritize smoothness, often fail to capture these dynamic structural features, particularly in high-dimensional settings. This article introduces Adaptive Joint Learning (AJL), a hierarchical regularization framework designed to integrate functional variable selection with structural changepoint detection in multivariate time-varying coefficient models. Unlike standard simultaneous estimation approaches, we propose a theoretically grounded two-stage screening-and-refinement procedure. This framework first synergizes adaptive group-wise penalization with sure screening principles to robustly identify active predictors, followed by a refined fused regularization step that effectively borrows strength across multiple outcomes to detect local regime shifts. We provide a rigorous theoretical analysis of the estimator in the ultra-high-dimensional regime (p » n). Crucially, we establish the sure screening consistency of the first stage, which serves as the foundation for proving that the refined estimator achieves the oracle property-performing as well as if the true active set and changepoint locations were known a priori. A key theoretical contribution is the explicit handling of approximation bias via undersmoothing conditions to ensure valid asymptotic inference. The proposed method is validated through comprehensive simulations and an application to Sleep-EDF data, revealing novel dynamic patterns in physiological states.

💡 Research Summary

The paper addresses the growing need to analyze high‑dimensional longitudinal data that exhibit both smooth time‑varying effects and abrupt regime shifts, a situation common in modern biomedical and econometric studies. Traditional functional data analysis (FDA) methods focus on smoothness and struggle with abrupt changes, while standard high‑dimensional regression techniques (e.g., Lasso, Group Lasso) lack the ability to detect structural breakpoints in functional coefficients. To fill this gap, the authors propose Adaptive Joint Learning (AJL), a hierarchical regularization framework that simultaneously performs functional variable selection across multiple correlated outcomes and detects changepoints in the intercept trajectories.

Model formulation
For subject i observed at times t_{il} with K outcomes, the time‑varying coefficient model is
( y_{ilk} = \alpha_k(t_{il}) + \sum_{j=1}^p x_{ij}(t_{il})\beta_{jk}(t_{il}) + \varepsilon_{ilk} ).
Both intercepts (\alpha_k(t)) and coefficient functions (\beta_{jk}(t)) are approximated by a common set of M B‑spline basis functions, turning the infinite‑dimensional problem into an ultra‑high‑dimensional linear regression with coefficient blocks (B_j\in\mathbb{R}^{M\times K}) and intercept vectors (a_k\in\mathbb{R}^M).

Penalty design

1. Adaptive Group Lasso on the Frobenius norm (|B_j|_F) encourages entire blocks to be zero, achieving functional variable selection across all outcomes. Adaptive weights are derived from an initial estimator, reducing bias and enabling oracle‑type performance.
2. Adaptive Fused Lasso on adjacent B‑spline coefficients of each intercept, (\sum_{m}|a_{k,m+1}-a_{k,m}|), enforces piecewise‑constant intercept trajectories, thereby detecting shared changepoints across outcomes.
3. The two penalties are combined in a single convex objective, but the underlying “ideal” problem is non‑convex. The authors employ a one‑step Local Linear Approximation (LLA) to approximate the non‑convex penalty, preserving computational tractability.

Two‑stage screening‑refinement
Stage 1: A sure‑screening procedure based on the adaptive group penalty reduces dimensionality from p to a manageable set while guaranteeing that, with probability tending to one, all truly active predictors are retained (sure screening consistency).
Stage 2: Using the reduced set, the refined estimator solves the fused‑penalty problem with LLA, achieving the oracle property—its asymptotic distribution matches that of an estimator that knows the true active set and changepoint locations a priori.

Theoretical contributions

• Uniform convergence rates that simultaneously handle the discrete changepoint set and the continuous functional space.
• Explicit treatment of approximation bias via an undersmoothing condition (M grows slowly, e.g., (M\asymp N^{1/5})), ensuring valid pointwise confidence bands.
• Proofs of non‑asymptotic error bounds, selection consistency, and asymptotic normality under ultra‑high‑dimensional scaling (p ≫ n).
• Novel incoherence conditions linking the B‑spline design matrix and the difference operator, required to separate smooth coefficient variation from true intercept jumps.

Algorithm
A block coordinate descent (BCD) combined with ADMM (BCD‑ADMM) solves the penalized least‑squares problem efficiently. Each block (either a (B_j) or an (a_k)) admits a closed‑form update after soft‑thresholding, and the overall scheme enjoys global convergence guarantees. Computational complexity per iteration is (O(NM(p+K))), making the method scalable to thousands of predictors and hundreds of time points.

Empirical evaluation
Simulation studies vary signal‑to‑noise ratios, number of changepoints, sparsity levels, and correlation structures among predictors. AJL consistently outperforms competing methods (standard Group Lasso, SCAD, MCP, fused Lasso) in terms of (i) true‑positive and false‑negative rates for variable selection, (ii) accuracy of estimated changepoint locations, and (iii) mean‑squared error of the functional estimates. Importantly, AJL retains sharp jumps without over‑smoothing, a limitation of traditional smoothness penalties.

Real data application
The authors apply AJL to the Sleep‑EDF database, modeling five EEG frequency‑band power trajectories (δ, θ, α, β, etc.) as multivariate outcomes. AJL uncovers shared changepoints that align with NREM‑REM transitions and provides pointwise confidence bands for the effects of age and gender, demonstrating the method’s capacity to handle high‑dimensional covariates (e.g., genetic risk scores) while respecting the complex temporal dynamics of sleep physiology.

Conclusion and outlook
Adaptive Joint Learning delivers a unified solution for (1) ultra‑high‑dimensional functional variable selection, (2) multi‑task learning across correlated outcomes, and (3) detection of abrupt structural changes, all while achieving oracle‑level statistical guarantees. The paper opens avenues for extensions to generalized link functions, time‑varying network structures, and Bayesian formulations, positioning AJL as a versatile tool for modern longitudinal data analysis.