LPClass: Lightweight Personalized Sensor Data Classification in Computational Social Systems

The blend of topics in computational social science enhances the research complexity in developing efficient computational social systems (CSSs). Electronic health (E-health) is a critical branch of CSS. Artificial intelligence-based cognitive computing is especially appropriate for solving E-health problems in social science. The development of the Internet of Things (IoT) and sensor technologies is triggering data explosion in E-health CSS. The IoT-based edge computing has been applied in the field of E-health to reduce the latency of data transmission. However, small edge devices have limited resources (e.g., computational and storage resources). There is an urgent need to develop lightweight and efficient classification models to classify E-health sensor data in edge computing. Automatic health sensor data classification can help medical workers make correct clinical decisions. Also, patient-specific modeling in E-health is important. Using personalized classification model can achieve higher diagnose accuracy than the generic models trained based on historical datasets. To address the above problems, we propose a lightweight personalized sensor data classification model, called LPClass. It embeds the shallow recurrent neural network as a kernel, which makes it lightweight enough to be deployed on edge devices. In addition, a transfer learning algorithm is proposed to build personalized models for individuals. We conduct comprehensive experiments to evaluate LPClass from different aspects. Compared with the generic models, the personalized models in LPClass can achieve a fast convergence rate while maintaining high classification accuracy.

Automated summary

The blend of topics in computational social science enhances research complexity in developing efficient computational social systems, with a focus on electronic health (E-health). The development of IoT-based edge computing in E-health faces challenges of limited resources and the need for lightweight and efficient classification models.