Neuro-Symbolic Attention Transformers for IoT Sensor Analysis: Integrating Symbolic Temporal Knowledge with Deep Learning
Year: 2026
Type: conference-paper
Abstract: Deep learning models based on transformers achieve excellent results across domains, yet they often ignore essential temporal knowledge embedded in Internet of Things sensor data. This paper presents a Neuro-Symbolic Attention Transformer that unites symbolic temporal knowledge with neural attention through a learnable fusion mechanism. The hybrid attention combines data-driven neural attention with rule-based symbolic attention via trainable weights that adapt to task complexity. Experiments on two contrasting Internet of Things datasets reveal when symbolic integration is most valuable. On a binary room occupancy task, the proposed model achieves competitive 99.15 percent accuracy without outperforming baselines, as strong discriminative signals reduce the need for symbolic augmentation. On the six-class Human Activity Recognition task, it reaches 92.11 percent accuracy, surpassing attention-based long short-term memory by 5.11 percentage points and standard transformers by 0.85 percentage points. Ablation studies and fusion weight analysis confirm that learnable integration is the key design choice, as the model down-weights symbolic cues on simple problems and relies on them more heavily as complexity increases. These findings offer evidence-based guidance on when symbolic knowledge engineering improves deep learning for Internet of Things sensor analysis.
Institutions: Mamun University, Urgench State University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7202300647