Univrsity Of Tehran Press Journal of Information Technology Management 2980-7972 18 4 2026 10 01 A Novel Edge AI Framework for Selective Subtask Offloading in UAV-Assisted Multiuser Multiserver Mobile Edge Computing 44 58 108914 10.22059/jitm.2026.108914 EN Suhel Ahamed Department of Information Technology, Guru Ghasidas Vishwavidyalaya (A Central University), Bilaspur, Chhattisgarh, India. Amit Kumar Dewangan Department of Information Technology, Guru Ghasidas Vishwavidyalaya (A Central University), Bilaspur, Chhattisgarh, India. Abhishek Jain Department of Information Technology, Guru Ghasidas Vishwavidyalaya (A Central University), Bilaspur, Chhattisgarh, India. Nishant Behar Department of Computer Science and Engineering, Guru Ghasidas Vishwavidyalaya (A Central University), Bilaspur, India. Pankaj Shankar Shrivastava Department of Electronics and Communication Engineering, Guru Ghasidas Vishwavidyalaya (A Central University), Bilaspur, India. Journal Article Mobile edge computing (MEC) integrated with unmanned aerial vehicles (UAVs) enables low-latency, energy-efficient computation for resource-constrained ground users; however, selective subtask offloading in multiserver, multiuser environments remains a challenging combinatorial optimization problem. A novel IT management framework is needed to address the challenges of modern smart computing environments and support efficient, adaptive operational decision-making. This paper proposes an Edge AI framework for selective subtask offloading built on the public IEEE Dataport MSMU-CO dataset; we use the term oracle to denote the minimum-cost multi-commodity flow (MCMF) solution, which serves as an offline near-optimal benchmark for generating training labels and evaluating policies. Starting from 800 MCMF-generated multi-server, multiuser scenarios, each user task is decomposed into K=3subtasks, oracle labels are derived under an explicit latency-energy cost model, and the data are augmented with UAV-supported MEC geometry, including aerial and ground servers, 3D positions, line-of-sight (LoS) probability, and effective uplink rates. To ensure realistic deployment conditions, a deliberate feature leakage guard eliminates anchor-derived geometries, preventing data leakage from post-optimization target assignments and maintaining causal inference realism. A compact multilayer perceptron (MLP) is then trained to predict per-subtask offloading decisions solely from task, channel, and localized UAV context features. Evaluated on a test set containing 1,780 users and 5,340 subtasks, the proposed model achieves an accuracy of 0.842, precision of 0.793, recall of 0.870, an F1 score of 0.830, and an AUC of 0.926 under the cost model used for label generation. Furthermore, we benchmark our approach against a greedy linear relaxation baseline (LR Greedy) based on Huang et al. (2018, 2020). Experiments show that both the proposed policy and the greedy baseline significantly reduce deadline violations and mean system costs compared with an All-Local execution strategy, effectively bridging the performance gap toward the offline oracle labels in varying UAV-assisted MEC environments. https://jitm.ut.ac.ir/article_108914_ae0e1b9353d3f43727c6bd6b902c8499.pdf