Balancing Performance and Energy Consumption of Bagging Ensembles for the Classification of Data Streams in Edge Computing

Guilherme Weigert Cassales, Heitor Murilo Gomes, Albert Bifet, Bernhard Pfahringer, Hermes Senger

Abstract

Edge computing facilitates low-latency data processing by bringing computation closer to IoT sources. However, the energy constraints of edge devices often conflict with the high computational demands of ensemble machine learning models. This paper presents a mini-batching strategy for parallel bagging ensembles designed to optimize the three-way trade-off between predictive accuracy, throughput, and energy consumption. Experimental results across diverse datasets and hardware platforms demonstrate that mini-batching reduces energy consumption in 96% of tested scenarios with negligible accuracy loss.

Introduction

The proliferation of Internet of Things (IoT) devices has led to the rise of Edge Computing (EC), where data streams are processed locally to ensure real-time responsiveness. While ensemble methods like Random Forests are highly effective for handling "concept drift" in non-stationary data streams, their energy footprint is often overlooked. This study addresses the "energy bill" of these models, focusing on making sophisticated classification algorithms viable for resource-constrained, battery-powered edge nodes.

Methodology: Parallel Bagging and Mini-batching

The researchers focus on Bagging Ensembles, where multiple base learners are trained independently. This structure is inherently parallelizable, as each learner can process data without waiting for others.

Bagging ensemble architecture for data streams.

The core innovation is the introduction of mini-batching into the online learning process. Traditionally, online learners process instances one by one as they arrive. By accumulating instances into a mini-batch (the authors found 50 instances to be optimal), the system can:

  1. Enhance Cache Locality: Grouped processing allows the CPU to reuse data already loaded into high-speed caches, reducing the frequency of power-intensive main memory accesses.
  2. Exploit Idle States: Between batch processing cycles, the system can utilize Dynamic Power Management (DPM) to put CPU cores into low-power sleep states, significantly reducing the overall energy draw.

Experimental Setup

The study evaluated six state-of-the-art algorithms, including Adaptive Random Forest (ARF) and Streaming Random Patches (SRP), implemented within the MOA (Massive Online Analysis) framework. These were tested against five diverse benchmark datasets: Airlines, GMSC, Electricity, Covertype, and Kyoto.

Hardware specifications and experimental measurement setup.

Testing spanned three hardware tiers to validate the strategy across the edge-to-cloud spectrum: a high-end dual Intel Xeon server, a personal computer (Intel i5), and a low-power single-board computer (Raspberry Pi). Energy consumption was measured using a dedicated sensor-logger setup to capture the real-time power draw of the processor during stream processing.

Key Findings and Results

  • Energy Efficiency: The mini-batching strategy reduced energy consumption in 96% of the experimental scenarios evaluated. The primary driver of these savings was the "nap" periods enabled between batches.
  • Throughput Gains: In many cases, throughput (instances processed per second) increased due to improved memory access patterns and reduced overhead.
  • Predictive Stability: The "cost" of batching was minimal; accuracy remained almost constant, typically staying within 1% of the non-batched baseline.
  • Hardware Sensitivity: The benefits were consistent across all tested platforms, proving that the strategy is robust regardless of the underlying architecture.

Conclusion and Future Work

Mini-batching provides a simple yet effective mechanism to balance the performance-energy trade-off in edge-based data stream classification. While the current study used static batch sizes, future research will explore adaptive mini-batching, where the batch size adjusts dynamically based on stream intensity and remaining battery life. Additionally, the authors aim to investigate how this strategy interacts with other hardware-specific optimizations like Dynamic Voltage and Frequency Scaling (DVFS).

Summary updated

BibTeX
@article{cassales2023balancing,
  author = {Guilherme Weigert Cassales and Heitor Murilo Gomes and Albert Bifet and Bernhard Pfahringer and Hermes Senger},
  title = {{Balancing Performance and Energy Consumption of Bagging Ensembles for the Classification of Data Streams in Edge Computing}},
  journal = {IEEE Trans. Netw. Serv. Manag.},
  year = {2023},
  doi = {10.1109/tnsm.2022.3226505},
}
Figure