A High-Resolution Dataset for Machine Learning-Based Link Adaptation in Urban NLOS Vehicular Networks

The advancement of machine learning (ML)-driven link adaptation (LA) in Vehicular Ad Hoc Networks (VANETs) is fundamentally constrained by the absence of high-fidelity, publicly available datasets containing synchronized physical-layer (PHY) and mobility measurements. While existing VANET datasets predominantly capture application-layer or macroscopic traffic metrics, they critically lack the per-packet channel state information (CSI), signal-to-noise ratio (SNR), Doppler dynamics, and interference characteristics essential for training robust, real-time LA models. To bridge this gap, we present VANET-LA, a novel high-resolution dataset explicitly engineered for ML-based link adaptation research in urban non-line-of-sight (NLOS) scenarios. The dataset is generated using a physics-informed methodology aligned with IEEE 802.11p standards, accurately capturing dominant impairments, including time-varying Doppler spread (0–1500 Hz), corresponding to relative speeds of up to approximately 275 km/h at a carrier frequency of 5.9 GHz; multipath delay dispersion (0–1 µs); inter-carrier interference (ICI); and outage-constrained fading. It also enforces conservative, mobility-aware modulation and coding scheme (MCS) selection under hysteresis and effective bandwidth constraints, reflecting real-world congestion control. The resulting dataset comprises 2,000 balanced samples spanning nine critical features: bandwidth, SNR, Doppler shift, delay spread, MCS index, modulation order, code rate, packet error rate (PER), and throughput. Deliberate class balancing is applied to eliminate representation bias inherent in raw simulation outputs. Rigorous validation confirms the dataset’s physical credibility through consistent relationships between channel impairments and link performance metrics under diverse urban NLOS mobility conditions. By providing synchronized, fine-grained PHY-layer measurements with explicit mobility context, VANET-LA establishes a reproducible foundation for developing, benchmarking, and deploying data-driven LA strategies that enhance reliability and spectral efficiency in safety-critical vehicular communications. Additionally, a benchmark evaluation using a Random Forest classifier achieved 85.3% accuracy in MCS prediction, demonstrating the dataset’s readiness for machine learning applications.