Skip to Content
SensorsSensors
  • This is an early access version, the complete PDF, HTML, and XML versions will be available soon.
  • Article
  • Open Access

29 September 2026

28 Pages

DSTGAT: A Spatio-Temporal Dynamic Graph Attention Network for User Activity Recognition

,
,
and
1
School of Information and Software Engineering, University of Electronic Science and Technology of China, Chengdu 610054, China
2
School of Computer Science and Software Engineering, Chengdu Jincheng College, Chengdu 611731, China
*
Author to whom correspondence should be addressed.
†
These authors contributed equally to this work.
This article belongs to the Section Intelligent Sensors

Abstract

Human Activity Recognition (HAR) in smart home environments is essential for Ambient Assisted Living (AAL). However, existing methods utilizing Graph Neural Networks (GNNs) typically rely on static sensor topologies and fail to adequately capture the dynamic spatio-temporal dependencies and heterogeneous nature of multi-sensor data. To address these challenges, this paper proposes a Dynamic Spatio-Temporal Graph Attention Network (DSTGAT) for robust HAR in smart homes. First, a multi-modal feature embedding mechanism is introduced to unify temporal, spatial, sensor type, and observation data into a cohesive representation. Second, a Cascade-LSTM architecture, combining bidirectional and unidirectional LSTMs, is designed to model the asymmetric temporal dependencies of human activities. Furthermore, to overcome the limitations of fixed graphs, a dynamic graph learning module based on Gumbel-Softmax sampling is proposed to adaptively reconstruct the sensor adjacency matrix. Finally, a Talking-Head graph attention mechanism is employed to facilitate cross-head feature interaction, enhancing global semantic fusion. Extensive experiments on four public CASAS smart home datasets (Aruba, Milan, Cairo, and Kyoto7) demonstrate that DSTGAT attains the highest F1 score on all four datasets, with improvements of 2.70 to 9.26 percentage points over the strongest baseline on each dataset; repeated runs with paired statistical tests confirm that these margins are significant on three of the four datasets (p < 0.01). Additionally, the proposed model degrades by no more than 0.6 percentage points with 30% missing sensor data on three of the four datasets. This research provides a highly efficient and reliable solution for non-intrusive activity monitoring in smart home deployments.

Article Metrics

Citations

Article Access Statistics

Multiple requests from the same IP address are counted as one view.