A Novel Temporal Leakage Evaluation Framework for Vehicle‐To‐Infrastructure Network Traffic Prediction

ABSTRACT Random train/test splitting is standard practice in V2I traffic prediction, yet it introduces a systematic flaw: When time‐series data are highly autocorrelated, randomly assigned test samples share temporal neighbors with the training set, causing models to appear more accurate than they are in deployment. This paper quantifies that inflation. We derive a synthetic vehicle‐to‐infrastructure (V2I) traffic dataset from the PKDD‐8 drone‐trajectory corpus using a long‐distance path loss model at 5.9 GHz and a speed‐dependent SNR decay function, producing 1938 half‐second bins across a 285‐m expressway segment. Three architectures, LSTM (Long Short‐Term Memory), BiLSTM‐MHA (Bidirectional Long Short‐Term Memory with Multi‐Head Attention), and CNN‐BiLSTM‐MHA (Convolutional Neural Network–Bidirectional Long Short‐Term Memory with Multi‐Head Attention), are evaluated under two conditions: a conventional random split (EXP‐1) and a strict chronological split (EXP‐2). We introduce the leakage gap, = (EXP‐1) − (EXP‐2), as a scalar diagnostic that can be computed from any pair of published values without rerunning experiments. Under strict chronological evaluation, the CNN‐BiLSTM‐MHA achieved the highest leakage‐free accuracy ( = 0.9558, five‐seed mean) and the smallest leakage gap ( = 0.0077, five‐seed mean). The BiLSTM‐MHA exhibited the largest single‐seed leakage gap ( = 0.00598) owing to its backward pass exploiting future temporal context available in randomized splits but absent in chronological evaluation, with a five‐seed mean of = 0.0100. The LSTM, despite yielding a low single‐seed leakage gap (Δ R 2 = 0.0115, σ = 0.00783), exhibits a ranking inversion that exposes its high initialization‐dependent leakage variance and directly demonstrates the necessity of multi‐seed evaluation. Multi‐seed Wilcoxon testing confirmed these gaps are statistically significant ( p = 0.03125). These results demonstrated that the prior state‐of‐the‐art of = 0.970 in this domain is likely inflated by temporal leakage. We propose the leakage gap metric as a practical tool for the ITS community to audit evaluation integrity in existing and future benchmarks.