Author(s)
Yanzhi Li
Abstract
Wireless networks are increasingly dense and heterogeneous, leading to complex interference patterns that challenge traditional resource allocation methods. In this work, we leverage the graph neural network (GNN) framework to model and optimize interference management as a graph coloring problem. By representing access points and user equipment as nodes and interference relationships as edges, we transform the interference avoidance task into a multicolor node classification problem. Inspired by the Potts model from statistical physics, we design an unsupervised loss function that penalizes conflicting assignments—discouraging adjacent nodes from sharing the same state. We demonstrate that this method can dynamically assign frequency channels or time slots with minimal conflicts, improving network throughput and reliability. This work offers a novel, scalable, and interpretable tool for next-generation network optimization.