Lisdatul Alifa, Irma Zakia, Hendrawan Hendrawan, Baso Maruddani Djaali, Kimi Rafif Asyadda, Fathoni Ubaidillah
Pilot contamination significantly degrades the minimum spectral efficiency (SE) in cell-free massive MIMO (CF-mMIMO) systems by inducing coherent interference. To mitigate this effect, effective pilot assignment and power allocation mechanisms are necessary. As such, strongly interfering user equipments (UEs) are prohibited in sharing the same pilot sequence, while remaining coherent interference is suppressed by controlling the transmit power. The existing heuristic or graph-free deep neural networks (DNNs) methods often fail to exploit the dependency between pilots and power, while suffering from limited generalization in larger networks. We design an unsupervised graph isomorphism network (GIN) with edge features for joint pilot assignment and power allocation (JPAPC) with the objective of maximizing the minimum SE of UEs in an uplink CF-mMIMO system. Inspired by graph coloring, we model the network as a homogeneous UE-UE graph, where UEs are defined as nodes, while edges capture pilot contamination level and geometric relationships between UE pairs. The proposed framework adopts sum aggregation and multilayer perceptron (MLP)-based updates to achieve injective message passing. This property enhances GIN's ability to distinguish complex interference structures that other graph neural network (GNN) architectures may fail to separate. This is important since different combinations of pilot-sharing UEs and overlapping access points (APs) may yield similar aggregate interference at the level of node features, but the underlying structures are fundamentally different. We compare the proposed GIN with state-of-the-art DNN and edge-weighted GNN (EW-GNN) baselines, where the latter also incorporates an edge-aware message passing but maps interference values to normalized weights during aggregation. Across different UE densities, we reveal that GIN consistently outperforms all baselines in terms of the worst-case and 95%-likely SE. Moreover, GIN shows robust generalization even when compared to EW-GNN, highlighting the effectiveness of its injective sum aggregration in preserving true interference structures. © 2020 IEEE.
Institut Teknologi Bandung, School Of Electrical Engineering And Informatics, Bandung, 40132, Indonesia; Universitas Negeri Jakarta (UNJ), Electrical Engineering Department, Jakarta, 13220, Indonesia
Research at a Glance
Register to unlock
Topics & SDG Alignment
Register to unlock
Collaboration
Register to unlock
Author Profile (Selected)
Register to unlock
References Overview
Register to unlock
Journal & Source
Register to unlock
Metadata & Integrity
Register to unlock