PER-SENSOR LSTM BASELINE AND SPARSE-COVERAGE ANALYSIS FOR SHORT-TERM URBAN TRAFFIC SPEED PREDICTION: A REPRODUCIBLE EVALUATION ON METR-LA

Short-term traffic speed prediction on the METR-LA benchmark is commonly evaluated by collapsing all 207 sensor readings into a single mean speed before prediction. This practice yields artificially low error metrics and produces results structurally incomparable with graph-based methods — an artefact that overstates the apparent advantage of graph architectures by a factor of approximately 2.83× at the 15-minute horizon. This paper makes two contributions: (1) a reproducibility and benchmark-hygiene contribution — a per-sensor LSTM architecture based on a shared-weight tensor-reshape strategy [B, N, T] → [B×N, T, 1], evaluated under the identical protocol used by DCRNN, Graph WaveNet, and MSTGACN, achieving MAE = 3.03 mph and RMSE = 5.73 mph at 15 minutes with 52,611 parameters; and (2) a sparse-coverage contribution — three experimental protocols applied to the trained baseline, revealing that block removal is the least damaging missing-data pattern for a spatial-information-free model, while sensor-level dropout is the most damaging — directly motivating a graph-convolutional successor for cities where detector coverage falls below 50%.

Hali tarjima qilinmagan