A Deep Multi-Task Learning Network for Activity Recognition and User Identification Using Smartphone Sensors
Sakorn Mekruksavanich, Wikanda Phaphan, Anuchit Jitpattanakul · Procedia Computer Science · 2025
Human activity recognition (HAR) and user identification utilizing smartphone sensors are crucial in domains such as wellness tracking, tailored services, and security. Conventional approaches handle these activities as distinct entities, leading to less-than-ideal performance and restricted applicability. This research presents a novel deep multi-task learning network that simultaneously performs activity recognition and user identification using smartphone sensor data. By leveraging shared representations and task-specific information, the network enhances accuracy. It combines a shared convolutional neural network (CNN) with bidirectional long short-term memory (BiLSTM) layers to extract features from raw sensor data. The CNN identifies patterns, while the BiLSTM captures sequential dependencies and contextual information. These features are processed by fully connected layers for both activity recognition and user identification. A multi-task loss function is proposed to optimize both tasks based on their complexity and importance. The approach is tested on the UCI-HAR dataset, containing diverse activities from multiple users. Experimental results show that the proposed CNN-BiLSTM model achieves 97.64% accuracy in activity recognition and 82.59% in user identification, surpassing current state-of-the-art methods. This demonstrates the effectiveness of the multi-task learning framework and the advantage of using the shared CNN-BiLSTM architecture for improving both tasks.