AKTMD: Attention-KAN-Based Neural Networks for Transportation Mode Detection
In recent years, with the development of sensor technology and the popularity of smartphones, transportation mode detection has become increasingly integrated into people’s lives, assisting various mobile intelligent services in improving personal health, optimizing urban traffic, etc. Al...
Saved in:
| Main Authors: | , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
IEEE
2025-01-01
|
| Series: | IEEE Access |
| Subjects: | |
| Online Access: | https://ieeexplore.ieee.org/document/10960293/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | In recent years, with the development of sensor technology and the popularity of smartphones, transportation mode detection has become increasingly integrated into people’s lives, assisting various mobile intelligent services in improving personal health, optimizing urban traffic, etc. Although previous studies have addressed fine-grained transportation mode detection, challenges including complex temporal correlations, non-linearities and interpretability required to further addressed. To this end, we propose a novel attention-KAN (Kolmogorov-Arnold Networks) based deep model for transportation mode detection, which primarily consists of three modules: convolutional KAN, multi-head attention and linear KAN. By introducing learnable B-spline functions, KAN can express more abundant feature information with fewer parameters which strengthens the non-linearities modeling and interpretability. Specifically, we employ convolutional KAN, which combines the advantages of KAN and convolutional neural network (CNN) to capture the short-term temporal correlations for distinct sensor features. Then we exploit multi-head attention to effectively capture long-term temporal correlations for linked sensor features in parallel, resulting in better trade-off between accuracy and efficiency. Extensive experimental results demonstrate that our proposed AKTMD model outperforms the state-of-the-art TMD-BERT model among the compared baselines, achieving an accuracy of 98.12% on the SHL dataset and 95.52% on the HTC dataset, with improvements of 5.05% and 3.29% over the TMD-BERT model on each dataset. |
|---|---|
| ISSN: | 2169-3536 |