Transportation mode detection through spatial attention-based transductive long short-term memory and off-policy feature selection
Mahsa Merikhipour, Shayan Khanmohammadidoustani, Mohammadamin Abbasi
Abstract
With mobile internet technology improving quickly, smartphones with many sensors have become increasingly popular for detecting transportation modes. Transportation modes are crucial for urban planning, managing traffic, and planning journeys. Although significant efforts have been made using traditional machine learning algorithms for this task, these methods often fall short because they rely on manually created features. Furthermore, while LSTM-based models are used for transportation mode detection, they often find it difficult to identify specific patterns in complex data over time and to choose the most useful features. To address these issues, we developed an advanced transportation mode detection model using a special kind of network called transductive long short-term memory (TLSTM) with spatial attention, improved further by an optimization algorithm known as off-policy proximal policy optimization (PPO) for better feature selection. The TLSTM can detect subtle changes in data over time, providing significant improvements over standard LSTM models by focusing on data points close to what will be tested. The off-policy PPO helps our model find complex patterns and connections that others often miss, improving its predictions. We also fine-tuned the hyperparameters of our model with an artificial bee colony (ABC) algorithm for optimal performance. Our model performed better than other current models after extensive testing on the Sussex Huawei locomotion (SHL), HTC, and United States-transport mode detection (US-TMD) datasets. It achieved F-measures of 92.323 %, 91.151 %, and 91.352 % on these datasets, demonstrating the effectiveness of combining TLSTM models with spatial attention and PPO algorithms in greatly improving the accuracy and reliability of transportation mode detection. The source code of the paper is publicly available at https://github.com/MahsaMerikhipour/TMD .