Machine Learning Benchmarking for V2X Link‑Quality Inference Across Communication Types

Accurate, transparent assessment of V2X link quality is important for communication-aware autonomy and for stress-testing vehicular connectivity across heterogeneous deployment contexts. This paper consolidates TiHAN-V2X measurements into a single analysis-ready corpus comprising V2V, V2I, I2V, and V2C links. The merged table contains 17{,}283 records spanning twelve scenario identifiers (S1--S7, I2V-S1--S2, V2I-S1--S3), plus V2C traces without scenario identifiers; the released processing pipeline standardizes metadata and explicitly annotates outliers rather than silently discarding them. To unify regression and classification evaluations under a common operational target, we define a Vehicle Quality Index (VQI), a bounded 0--100 score that aggregates signal quality, throughput, delay, and packet error behavior after robust normalization. Using this target, we benchmark a set of classical and tree-based models under stratified holdout and leave-one-scenario-out validation, and we interpret fitted models via feature importance, SHAP explanations, partial dependence, and ablation analyses. Across evaluation settings, tree ensembles provide the strongest performance and yield consistent explanatory patterns, with packet error dynamics, latency variability, and SNR emerging as dominant determinants of VQI. All data artifacts and the end-to-end evaluation pipeline are released to support direct replication and to provide a baseline for scenario-aware V2X link-quality benchmarking.