SparseBonsai: A Dynamic, Resource-Efficient Classification Model for Edge Computing and Industrial IoT

Abstract

The rapid evolution of edge computing and IIoT ecosystems demands lightweight machine learning models that deliver accurate predictions under resource constraints. Traditional classifiers, such as deep neural networks and decision trees, often struggle to balance accuracy, interpretability, and computational efficiency in such environments. This work introduces SparseBonsai, an enhanced variant of the Bonsai Tree algorithm that includes projection techniques, dynamic sparsity parameters, and adaptive regularization. Bonsai Tree models work with fixed parameters, whereas SparseBonsai dynamically adjusts sparsity and regularization during training. It improves adaptability and generalization. SparseBonsai achieves 86.91% classification accuracy with a model size of 0.048 MB, and inference time is less than 35% as of neural networks with a competitive accuracy. Model’s robustness and efficiency is evaluated using precision, recall, F1-score, and ROC-AUC. These results show that the SparseBonsai can be a practical solution for a real-time and resource-efficient fault detection system as an IIoT edge computing platform. The novelty of SparseBonsai is its dynamic adjustment of sparsity and regularization in training in comparison to the conventional Bonsai Tree algorithm. SparseBonsai reduces inference time by 35% and memory footprint by 38% compared with existing lightweight classifiers.