Year: 2026
Type: preprint
Abstract: Sensor rankings can guide the restoration of disrupted traffic networks, but their value depends on the forecasts produced after selected readings are returned. We evaluate this decision by selecting sources on earlier observations and testing their joint restoration under shared synthetic faults. On PEMS03, the same Linear proxy selections that outperformed variance under Graph WaveNet increased mean absolute error by 0.330 vehicles per five minutes under a diffusion convolutional recurrent neural network (DCRNN). This prespecified primary contrast reversed in all three training seeds, five masks and weekly summaries. The DCRNN protocol was fixed after the Graph WaveNet results and before final output inspection. Rolling PEMS08 retained a Linear advantage under both predictors, with smaller differences against redundancy-aware alternatives. On METR-LA, the mean Linear advantage over variance was 0.000260 mph across three DCRNN seeds, with a conditional interval spanning zero; a separate long short-term memory reference also showed little separation. Exact-deletion and correlation-based policies produced further predictor-dependent outcomes. These results support evaluating source sets under the intended predictor and distinguishing the overall restoration benefit from the incremental benefit of their ranking.
Source: Research Square
Authors: Bailin Xiong, Kai Du, Jinhui Tang, Renxiang Wang
Institution: Chongqing Jiaotong University
Language: English
Subfield: Transportation
Field: Social Sciences
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W7225990338
