Learning and Reconstructing Conflicts in O-RAN: A Graph Neural Network Approach

TL;DR

This paper addresses the challenge of unseen conflicts among xApps in O-RAN Near-RT RIC by proposing a data-driven method that uses GraphSAGE to reconstruct conflict graphs from observed xApp, parameter, and KPI data. It introduces temporal graph representations and an MSE-based training objective to produce embeddings that reveal latent relationships, enabling adjacency reconstruction and threshold-based graph inference. Additionally, it defines graph-based criteria for direct, indirect, and implicit conflicts and demonstrates, on a literature-based conflict model, that the approach achieves high reconstruction accuracy and effective conflict labeling, especially for implicit and indirect conflicts with sufficient data and training. The work advances practical conflict management in O-RAN by enabling data-driven discovery of hidden dependencies and provides a pathway toward causal understanding and scalable mitigation.

Abstract

The Open Radio Access Network (O-RAN) architecture enables the deployment of third-party applications on the RAN Intelligent Controllers (RICs). However, the operation of third-party applications in the Near Real-Time RIC (Near-RT RIC), known as xApps, may result in conflicting interactions. Each xApp can independently modify the same control parameters to achieve distinct outcomes, which has the potential to cause performance degradation and network instability. The current conflict detection and mitigation solutions in the literature assume that all conflicts are known a priori, which does not always hold due to complex and often hidden relationships between control parameters and Key Performance Indicators (KPIs). In this paper, we introduce the first data-driven method for reconstructing and labeling conflict graphs in O-RAN. Specifically, we leverage GraphSAGE, an inductive learning framework, to dynamically learn the hidden relationships between xApps, parameters, and KPIs. Our numerical results, based on a conflict model used in the O-RAN conflict management literature, demonstrate that our proposed method can effectively reconstruct conflict graphs and identify the conflicts defined by the O-RAN Alliance.

Figures (10)

• Figure 1: Conflict graph illustrating potential known conflicts between xApps and their control parameters (identifiable before deployment), and unknown conflicts that may arise due to the non-trivial relationships between control parameters and KPIs.
• Figure 2: Illustration of the types of conflicts considered by the O-RAN Alliance using our graph labeling definitions, allowing us to identify conflicts in the structure of a conflict graph $\mathcal{G}$.
• Figure 3: Structure of the conflict model Banerjee2022TowardNetworks we used to validate our solutions for graph reconstruction and conflict labeling.
• Figure 4: Evolution of the correlation matrix of our GNN model with the number of epochs (created with a dataset size of 450 samples), illustrating the convergence.
• Figure 5: Conflict graph reconstruction accuracy according to the number of epochs and dataset size for a fixed threshold of 0.5.