DSHformer: Locality-Sensitive Hash Attention and Prototype Alignment for Sensor-Based Human Activity Recognition

Area of Science:

• Computer Science
• Machine Learning
• Signal Processing

Background:

• Sensor-based human activity recognition (HAR) is crucial for healthcare and wearables.
• Deep learning advances HAR, but distribution shifts and high computational complexity limit real-world use.
• Existing methods struggle with generalization across users/sensors and efficient long-sequence processing.

Purpose of the Study:

• To propose DSHformer, an accuracy-oriented HAR framework addressing generalization and efficiency.
• To combine compact channel-temporal encoding with locality-sensitive hashing (LSH)-based attention.
• To enable efficient and accurate HAR on resource-constrained wearable devices.

Main Methods:

• A low-parameter patch-based graph-attention encoder models sensor channel-temporal dynamics.
• A trainable prototype pool and decomposition network improve class separability via prototype alignment.
• A decomposition-stable LSH-based attention mechanism with O(LlogL) complexity is introduced for HAR.

Main Results:

• DSHformer achieved high accuracies on five benchmarks: WISDM (98.6%), UCI-HAR (93.7%), PAMAP2 (98.4%), Opportunity (88.5%), and UniMiB-SHAR (96.6%).
• The framework demonstrated competitive or superior performance against Transformer variants and HAR-specific baselines.
• Ablation studies confirmed the significant contribution of each DSHformer component.

Conclusions:

• DSHformer effectively addresses distribution shift and computational complexity in sensor-based HAR.
• The proposed LSH-based attention mechanism enables efficient long-sequence modeling.
• DSHformer offers a promising solution for real-world, resource-constrained HAR applications.