A GNN-Based Discrete Choice Experiment for Heterogeneous Computational Task Offloading

A multitude of 6G smart devices will integrate applications that require contextual data processing, such as healthcare remote sensing and interactive applications. These are computationally intensive tasks that require quick execution within stringent time constraints. Given the heterogeneity of these applications as well as the communication and computation uncertainty due to the shared nature of the edge environment, computation competition and access interference pose serious challenges in the offloading of computational tasks. In this respect, the offloading decision-making of edge devices is influenced and shaped by prospect-theoretic characteristics. To address this challenge, a graph neural network (GNN)-based discrete choice experiment (DCE) is proposed to model the choice of the appropriate offloading server based on a feature — attribute matching. An encoded vector that is validated using interference, channel gain, and transmission rates, is encoded into the embedding space, and offloading solution for each edge device is obtained. The performance evaluation results show that the proposed GNN-based DCE lowers the expected overheads, which maximizes the amount of offloaded computational tasks.