Data and channel-aware device selection for split learning in resource-constrained B5G/6 G networks

Abstract

Abstract Split learning (SL) is a promising distributed learning approach for fifth-generation/sixth-generation (B5G/6 G) networks. However, its performance is affected by bandwidth constraints and statistical heterogeneity in the data. Prior works have considered either channel-aware or data-aware device selection in isolation, but few jointly consider both under per-round participation limits. This article extends the “A Network Metric-Aware Split Learning Architecture for B5G/6 G Mobile Networks” (SLArch) framework with explicit bandwidth constraints, partial user equipment (UE) participation, and non-independent and identically distributed (non-IID) data modelled via Dirichlet distributions. We introduce four device selection policies: random, channel-quality-based (ChQual), data-quality-based (DataQual), and a joint channel-and-data-aware policy (ChDataQual) weighted by parameters $$\varphi $$ <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"> <mml:mi>φ</mml:mi> </mml:math> and $$\beta $$ <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"> <mml:mi>β</mml:mi> </mml:math> . Simulations with a total of 150 UEs and 10 participants per round demonstrate key network and SL performance results. DataQual achieves the highest accuracy across most slices, reaching 94.8% on Modified National Institute of Standards and Technology (MNIST) under massive machine-type communications (mMTC), 92.5% on Fashion-MNIST under enhanced mobile broadband (eMBB), and 86% on Federated Extended MNIST (FEMNIST) under eMBB. ChDataQual achieves 93.2% on Fashion-MNIST under eMBB, outperforming DataQual by 0.7%. In terms of reliability, ChQual selection achieves the lowest packet loss of 0.059% under ultra-reliable low-latency communications (URLLC). ChDataQual reaches 91.4% accuracy on MNIST under URLLC while reducing packet loss by 58% relative to DataQual. Furthermore, ChQual and ChDataQual exhibit lower and more stable energy consumption and jitter than Random and DataQual. Throughput remains robust across all strategies, showing that integrated channel-data awareness enables reliable, high-performance SL in resource-constrained B5G/6 G environments.