A Benchmark of Data Stream Classification for Human Activity Recognition on Connected Objects.

Area of Science:

• Computer Science
• Machine Learning
• Internet of Things (IoT)

Background:

• Connected devices increasingly rely on efficient data stream classification for applications like Human Activity Recognition (HAR).
• Evaluating stream classifiers requires assessing both accuracy and resource constraints inherent in edge computing environments.

Purpose of the Study:

• To evaluate the performance and resource consumption of five common data stream classification algorithms for HAR on connected devices.
• To compare stream classifiers against offline methods and identify challenges for future research in resource-constrained environments.

Main Methods:

• Implemented five stream classification algorithms (Hoeffding Tree, Mondrian Forest, Naïve Bayes, Feedforward Neural Network, Micro Cluster Nearest Neighbor) within a unified library.
• Tested algorithms on two real-world and three synthetic HAR datasets, measuring classification accuracy, runtime, memory usage, and power consumption.
• Assessed classifiers' ability to recover from concept drift.

Main Results:

• Hoeffding Tree, Mondrian Forest, and Naïve Bayes generally outperformed Feedforward Neural Network and Micro Cluster Nearest Neighbor on four of six datasets.
• Hoeffding Tree and Micro Cluster Nearest Neighbor demonstrated resilience to concept drift.
• All evaluated stream classifiers performed significantly worse than offline classifiers on real-world HAR data.
• Hoeffding Tree and Mondrian Forest exhibited high memory usage and runtime, while power consumption was consistent across algorithms.

Conclusions:

• Stream learning for HAR on connected devices faces significant challenges, notably high memory consumption and suboptimal F1 scores.
• Further research is needed to address these limitations for practical deployment on edge devices.