SMORE: Similarity-based Hyperdimensional Domain Adaptation for Multi-Sensor Time Series Classification

TL;DR

SMORE introduces a novel hyperdimensional computing-based domain adaptation framework for multi-sensor time series classification that explicitly accounts for domain context during inference. By encoding data into high-dimensional hypervectors, training domain-specific models, and constructing expressive domain descriptors, SMORE detects out-of-distribution samples and dynamically ensembles domain models at test time to tailor predictions. Empirical results on DSADS, USC-HAD, and PAMAP2 show SMORE achieving about 1.98% higher accuracy than state-of-the-art CNN-based DA methods, with substantial training and inference speedups (e.g., up to ~$19\times$ faster training and ~4.6x faster inference) and favorable performance on resource-constrained devices. The approach demonstrates strong robustness to distribution shifts while maintaining efficiency, making it well-suited for edge deployments in IoT environments.

Abstract

Many real-world applications of the Internet of Things (IoT) employ machine learning (ML) algorithms to analyze time series information collected by interconnected sensors. However, distribution shift, a fundamental challenge in data-driven ML, arises when a model is deployed on a data distribution different from the training data and can substantially degrade model performance. Additionally, increasingly sophisticated deep neural networks (DNNs) are required to capture intricate spatial and temporal dependencies in multi-sensor time series data, often exceeding the capabilities of today's edge devices. In this paper, we propose SMORE, a novel resource-efficient domain adaptation (DA) algorithm for multi-sensor time series classification, leveraging the efficient and parallel operations of hyperdimensional computing. SMORE dynamically customizes test-time models with explicit consideration of the domain context of each sample to mitigate the negative impacts of domain shifts. Our evaluation on a variety of multi-sensor time series classification tasks shows that SMORE achieves on average 1.98% higher accuracy than state-of-the-art (SOTA) DNN-based DA algorithms with 18.81x faster training and 4.63x faster inference.

Figures (7)

• Figure 1: Motivation of Our Proposed $\textit{$\mathsf{SMORE}$\xspace}$
• Figure 2: The Workflow of Our Proposed $\textit{$\mathsf{SMORE}$\xspace}$
• Figure 3: HDC Encoding for Multi-Sensor Time Series Data
• Figure 4: Comparing LODO Accuracy of $\textit{$\mathsf{SMORE}$\xspace}$ and CNN-based Domain Adaptation Algorithms
• Figure 5: Imapct of $\delta^*$ on Model Performance