A novel Heterogeneous Temporal Graph Neural Network framework has demonstrated a 42.4 percent reduction in mean absolute error and achieved an area under the receiver operating characteristic curve of 0.899 when predicting co-location interference risk scores and Service Level Agreement violation events in cloud computing environments.
Cloud data centers encounter persistent operational difficulties stemming from co-location interference. This phenomenon emerges whenever multiple virtual machines assigned to the same physical host machine compete directly for shared computing infrastructure, including central processing unit resources, system memory, and input and output bandwidth.
Conventional virtual machine placement strategies typically rely on standard bin-packing heuristics alongside their multi-objective or machine learning extensions. However, because these established strategies optimize workloads primarily during initial provisioning time, they fail to capture the dynamic, time-varying interactions that occur between competing applications across active servers.
To address both dynamic placement and migration scheduling, researchers developed the Heterogeneous Temporal Graph Neural Network architecture. The system models physical data center infrastructure as a heterogeneous graph consisting of virtual machine nodes, host nodes, runs-on relationship edges, and co-location edges connecting co-hosted instances.
The underlying neural architecture integrates a bidirectional Graph Attention Network with a Long Short-Term Memory encoder. Combining these two structures allows the framework to jointly track topological dependencies across the heterogeneous infrastructure graph and dynamic temporal workload patterns evolving over operating intervals.
The research team trained the multi-task model on empirically calibrated synthetic workload traces derived from two major industry sources: the Azure Virtual Machine Noise dataset and the 2023 Alibaba GPU Cluster Trace. This multi-task configuration enables the system to concurrently estimate interference risk scores and forecast impending Service Level Agreement violations.
In comparative experimental assessments, the framework’s area under the receiver operating characteristic curve reached 0.899, outperforming a Gradient Boosting Trees baseline that registered 0.711 and an isolated Long Short-Term Memory baseline that scored 0.763. Ablation testing confirmed that the Long Short-Term Memory temporal encoder and the co-location graph edges represent the model’s most impactful design components.
When evaluated under matched exogenous workload conditions, the framework’s proactive migration policy also produced higher physical resource efficiency. The proactive scheduling approach achieved 6.8 percent higher physical host utilization than a random-placement baseline and 3.5 percent higher utilization than a reactive threshold-based baseline.
Published in the International Journal of Networked and Distributed Computing on November 13, 2026, the paper introduces HT-GNN to resolve heuristic limitations from Section 2.1. Shared early to accelerate access to peer-reviewed findings, the citable manuscript carries a permanent digital object identifier and remains subject to editorial revisions before final Version of Record publication.
Subject of Research: Interference-aware virtual machine placement and proactive migration using heterogeneous temporal graph neural networks
Article Title: Interference Aware Virtual Machine Placement and Proactive Migration Using Heterogeneous Temporal Graph Neural Networks
Article References: Pote, S. A., Gupta, M., Kusane, S., Kumar, T., Sharma, G., Hemalatha, P. K., Kathuria, S., & Kochher, V. (2026). Interference Aware Virtual Machine Placement and Proactive Migration Using Heterogeneous Temporal Graph Neural Networks. International Journal of Networked and Distributed Computing. https://doi.org/10.1007/s44227-026-00138-8
Image Credits: AI Generated
DOI: 10.1007/s44227-026-00138-8
Keywords: virtual machine placement, proactive migration, graph neural networks, heterogeneous graphs, co-location interference, cloud data centers, graph attention networks, long short-term memory, service level agreements, Azure VM Noise dataset, Alibaba GPU Cluster Trace, gradient boosting trees, host utilization, mean absolute error, AUC-ROC, ablation experiments, bin-packing heuristics, workload interactions
Cite Scienmag News
Cassandra Pierce. (November 17, 2026). Temporal Graph Neural Network Cuts Virtual Machine Interference Prediction Error by 42.4%. Scienmag. https://scienmag.com/temporal-graph-neural-network-cuts-virtual-machine-interference-prediction-error-by-42-4/
Cassandra Pierce. "Temporal Graph Neural Network Cuts Virtual Machine Interference Prediction Error by 42.4%." Scienmag, 17 November 2026, https://scienmag.com/temporal-graph-neural-network-cuts-virtual-machine-interference-prediction-error-by-42-4/. Accessed 17 November 2026.
Cassandra Pierce. "Temporal Graph Neural Network Cuts Virtual Machine Interference Prediction Error by 42.4%." Scienmag. November 17, 2026. https://scienmag.com/temporal-graph-neural-network-cuts-virtual-machine-interference-prediction-error-by-42-4/








