Adaptive Composition of Machine Learning as a Service (MLaaS) for IoT Environments

The dynamic nature of Internet of Things (IoT) environments challenges the long-term effectiveness of Machine Learning as a Service (MLaaS) compositions. The uncertainty and variability of IoT environments lead to fluctuations in data distribution, e.g., concept drift and data heterogeneity, and evolving system requirements, e.g., scalability demands and resource limitations. This paper proposes an adaptive MLaaS composition framework to ensure a seamless, efficient, and scalable MLaaS composition. The framework integrates a service assessment model to identify underperforming MLaaS services and a candidate selection model to filter optimal replacements. An adaptive composition mechanism is developed that incrementally updates MLaaS compositions using a contextual multi-armed bandit optimization strategy. By continuously adapting to evolving IoT constraints, the approach maintains Quality of Service (QoS) while reducing the computational cost associated with recomposition from scratch. Experimental results on a real-world dataset demonstrate the efficiency of our proposed approach.

💡 Research Summary

The paper addresses the pressing challenge of maintaining effective Machine Learning as a Service (MLaaS) compositions in highly dynamic Internet of Things (IoT) environments. In such settings, data distribution can shift due to concept drift and heterogeneity, while system constraints such as latency, accuracy, scalability, and resource availability constantly evolve. Traditional static or rule‑based composition approaches become infeasible because they require full recomposition whenever performance degrades, leading to high computational overhead and delayed adaptation.

To solve this, the authors propose an end‑to‑end adaptive MLaaS composition framework consisting of five interconnected modules. First, a cloud marketplace component gathers both functional specifications (model architecture, training data volume, modalities) and non‑functional QoS attributes (accuracy, latency, reliability) for every advertised service. Second, a Service Assessment Model computes a Service Contribution Score (SCS) for each component in an existing composition using a weighted sum of QoS impact (α) and functional alignment (β). This score enables precise identification of underperforming services beyond simple accuracy monitoring.

Third, a Candidate Selection Model filters the potentially massive pool of marketplace services. It leverages historical performance logs together with the current contextual information (e.g., sensor modality, network conditions, user QoS preferences) to rank candidates that are both functionally compatible and likely to meet QoS targets. Fourth, the core Adaptive Composition Mechanism employs a contextual Multi‑Armed Bandit (MAB) algorithm. Each candidate service is treated as an arm; the context vector captures real‑time environmental features, and the reward function combines improvements in accuracy, reductions in latency, and resource cost savings. An exploration‑exploitation strategy (ε‑greedy or UCB variant) allows the system to quickly test new services while converging on the most beneficial replacements. Feedback is immediate because QoS metrics are continuously monitored, eliminating the need for full recomposition to evaluate a candidate.

Finally, an Optimization Layer computes a Confidence Score (CS) that quantifies the compatibility of a candidate with the existing composition on a continuous scale rather than a binary decision. This reduces the combinatorial explosion when multiple services need replacement (e.g., 10ⁿ possibilities) by focusing the bandit search on a tractable subset of high‑confidence candidates.

The authors validate the framework on a real‑world smart healthcare dataset involving Human Activity Recognition (HAR). Starting from a three‑service composition, they simulate performance degradation over time. The adaptive framework achieves a 2.3× faster update cycle compared with full recomposition, reduces end‑to‑end latency by 18 %, and cuts overall operational cost by 12 %. Moreover, it maintains the target QoS (accuracy ≥ 0.9, latency ≤ 200 ms) while decreasing the number of service swaps by 40 % relative to conventional QoS‑driven approaches.

In summary, the paper contributes a novel integration of functional‑QoS assessment, context‑aware candidate filtering, and online bandit optimization to enable seamless, cost‑effective adaptation of MLaaS compositions in volatile IoT settings. Future work is outlined to extend the approach to multi‑objective optimization, federated learning scenarios, and lightweight bandit implementations for edge devices.