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.
Introduction
In the fast-moving world of real-time machine learning, "concept drift"—the phenomenon where data distributions shift over time—is an ever-present challenge. If a model is not updated, its performance inevitably withers as the gap between its training distribution and the current environment grows. However, in resource-constrained environments such as mobile phones, IoT sensors, or edge gateways, updating a model too frequently can rapidly deplete battery life, consume expensive compute cycles, or exceed monetary budgets.
Traditional approaches to this dilemma typically fall into two categories: heavyweight statistical drift detectors (like ADWIN) which monitor raw data changes, or simple heuristics (like periodic updates). The former often consumes more power than the inference task itself, while the latter lacks theoretical guarantees for budget adherence or accuracy optimization.
Problem Formulation and Methodology
RCCDA (Resource-Constrained Concept Drift Adaptation) addresses this gap by framing model updates as a constrained optimization problem. The goal is to maximize long-term average accuracy while strictly adhering to a predefined resource budget.
The Lyapunov Drift-Plus-Penalty Framework
The core of RCCDA is the Lyapunov drift-plus-penalty framework, a control theory technique repurposed for ML training dynamics. The framework utilizes a virtual queue $Q(t)$ that acts as a resource debt accumulator:
- Accumulation: Every time the policy decides to update the model, the queue grows by the associated resource cost.
- Drainage: At every time step, the queue "drains" by the amount of the allowed average budget.
- Feedback: A large queue indicates that the budget has been overspent, increasing the "effective cost" of future updates in the decision-making process.
The Decision Rule
At each time step $t$, RCCDA evaluates a greedy-optimal decision rule. It triggers an update only if the estimated performance gain (the expected drop in loss) outweighs the effective resource cost, which is the actual cost weighted by the current state of the virtual queue. Crucially, RCCDA is "distribution ignorant"; it does not require complex statistical tests on raw input data. Instead, it uses a causal estimator based on historical inference loss—a signal readily available during standard model operation.
Experimental Evaluation
The researchers evaluated RCCDA across four diverse domain generalization datasets, simulating various drift schedules including Burst (sudden), Step (persistent), Wave (cyclic), and Spikes (transient):
- PACS: Images across Photo, Art, Cartoon, and Sketch domains.
- DigitsDG: Handwritten digits from multiple sources (MNIST, SVHN, etc.).
- OfficeHome: 65 object categories across four distinct image styles.
- MEMD-ABSA: Multi-domain textual sentiment analysis.
Results and Analysis
RCCDA demonstrates significant improvements over standard baselines like uniform random or periodic update policies.
Accuracy and Budget Stability
On the PACS dataset under "Burst" drift, RCCDA achieved 72.8% accuracy, significantly higher than the ~64-65% achieved by standard baselines. Furthermore, the paper provides theoretical proof that RCCDA’s constraint violation vanishes as $T \to \infty$, ensuring that long-term budgets are strictly respected.
Computational Efficiency
One of the most striking results is the negligible overhead of the RCCDA policy. The decision-making process takes approximately 0.7 microseconds. In contrast, the popular ADWIN drift detector is roughly 2,480 times slower, and the Kolmogorov-Smirnov (KS) test is over 169 million times slower. This efficiency makes RCCDA uniquely suited for edge deployment where the "observer effect" of monitoring drift must be minimized.
Robustness
The policy also exhibits high robustness to noise. When the loss signal was injected with 30% synthetic noise, RCCDA's accuracy dropped by only 2.5%, indicating its reliability in real-world scenarios where loss measurements may be imperfect.
Conclusion and Future Work
RCCDA provides a theoretically grounded, lightweight solution for managing the trade-off between model accuracy and resource consumption under concept drift. While highly effective, the policy's performance is tied to the accuracy of its performance gain estimator. Future work may explore deeper integration with Continual Learning (CL) techniques to address "catastrophic forgetting" during cyclic drift and automated hyperparameter tuning to further simplify deployment across diverse tasks.
Summary updated