Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

RCCDA: Adaptive Model Updates in the Presence of Concept Drift under a Constrained Resource Budget

Adam Piaseczny, Md Kamran Chowdhury Shisher, Shiqiang Wang, Christopher G. Brinton

TL;DR

RCCDA addresses real-time model adaptation under concept drift with strict resource budgets by deriving a Lyapunov drift-plus-penalty-based threshold policy that decides updates using only past loss information. The method provides provable bounds on both convergence (time-averaged gradient norms) and budget compliance (update-frequency constraint) while remaining lightweight and distribution-ignorant. The authors validate RCCDA on four domain-generalization datasets under multiple drift schedules, showing improved accuracy and faster drift recovery compared with baseline policies, all within the specified resource budget. The approach is practical for edge and real-time deployments, offering a principled, low-overhead alternative to drift-detection-based schemes and highlighting trade-offs between update frequency, drift, and performance.

Abstract

Machine learning (ML) algorithms deployed in real-world environments are often faced with the challenge of adapting models to concept drift, where the task data distributions are shifting over time. The problem becomes even more difficult when model performance must be maintained under adherence to strict resource constraints. Existing solutions often depend on drift-detection methods that produce high computational overhead for resource-constrained environments, and fail to provide strict guarantees on resource usage or theoretical performance assurances. To address these shortcomings, we propose RCCDA: a dynamic model update policy that optimizes ML training dynamics while ensuring compliance to predefined resource constraints, utilizing only past loss information and a tunable drift threshold. In developing our policy, we analytically characterize the evolution of model loss under concept drift with arbitrary training update decisions. Integrating these results into a Lyapunov drift-plus-penalty framework produces a lightweight greedy-optimal policy that provably limits update frequency and cost. Experimental results on four domain generalization datasets demonstrate that our policy outperforms baseline methods in inference accuracy while adhering to strict resource constraints under several schedules of concept drift, making our solution uniquely suited for real-time ML deployments.

RCCDA: Adaptive Model Updates in the Presence of Concept Drift under a Constrained Resource Budget

TL;DR

RCCDA addresses real-time model adaptation under concept drift with strict resource budgets by deriving a Lyapunov drift-plus-penalty-based threshold policy that decides updates using only past loss information. The method provides provable bounds on both convergence (time-averaged gradient norms) and budget compliance (update-frequency constraint) while remaining lightweight and distribution-ignorant. The authors validate RCCDA on four domain-generalization datasets under multiple drift schedules, showing improved accuracy and faster drift recovery compared with baseline policies, all within the specified resource budget. The approach is practical for edge and real-time deployments, offering a principled, low-overhead alternative to drift-detection-based schemes and highlighting trade-offs between update frequency, drift, and performance.

Abstract

Machine learning (ML) algorithms deployed in real-world environments are often faced with the challenge of adapting models to concept drift, where the task data distributions are shifting over time. The problem becomes even more difficult when model performance must be maintained under adherence to strict resource constraints. Existing solutions often depend on drift-detection methods that produce high computational overhead for resource-constrained environments, and fail to provide strict guarantees on resource usage or theoretical performance assurances. To address these shortcomings, we propose RCCDA: a dynamic model update policy that optimizes ML training dynamics while ensuring compliance to predefined resource constraints, utilizing only past loss information and a tunable drift threshold. In developing our policy, we analytically characterize the evolution of model loss under concept drift with arbitrary training update decisions. Integrating these results into a Lyapunov drift-plus-penalty framework produces a lightweight greedy-optimal policy that provably limits update frequency and cost. Experimental results on four domain generalization datasets demonstrate that our policy outperforms baseline methods in inference accuracy while adhering to strict resource constraints under several schedules of concept drift, making our solution uniquely suited for real-time ML deployments.

Paper Structure

This paper contains 37 sections, 3 theorems, 48 equations, 4 figures, 15 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Methodology
  4. System Model
  5. Problem Formulation
  6. Policy Design
  7. Policy Derivation
  8. Our Algorithm
  9. Theoretical Analysis
  10. Convergence Analysis
  11. Stability Analysis
  12. Experiments
  13. Conclusion and Limitations
  14. Societal Impact
  15. Convergence Proofs
  16. ...and 22 more sections

Key Result

Theorem 5.1

If Assumptions 1-2 hold and the learning rate is chosen such that $\eta < \frac{2 P^{\pi}_{\min}}{L}$, then the time-averaged gradient satisfies where $T$ is the total number of time steps, $\Delta f_{\delta}(t)=f(\theta_t, \mathcal{D}_{t+1}) - f(\theta_t, \mathcal{D}_t)$ is the time-varying drift-induced loss change, $\mu=\left(\eta P^{\pi}_{\min}-\frac{L\eta^2}{2}\right)$, and $P^{\pi}_{\min}$

Figures (4)

  • Figure 1: Interplay between policy design, concept drift, and model performance. (a) Optimal model point movement over time due to drift. Different policies result in varying total distances to the drifting optimal model. (b) Update timing influences average performance given set number of updates.
  • Figure 2: Dynamics of model accuracy and update rate for proposed and baseline policies under various concept drift schedules, with corresponding dataset composition over time. We see that the proposed policy recovers the quickest after sharp transitions while adhering to resource constraints.
  • Figure 3: Average validation accuracy of policies across different update rates for the burst drift schedule, across PACS and DigitsDG datasets. The left column is the accuracy per real update rate, and the right column is the accuracy per constrained update rate $\frac{\bar{\lambda}}{\lambda}$.
  • Figure : RCCDA

Theorems & Definitions (3)

  • Theorem 5.1
  • Corollary 5.2
  • Theorem 5.3