Sparse multimodal observability under degraded traffic sensing: critical-density theory and tail-risk evidence
Authors:
- Claire Y.T. ChenCCClaire Y.T. Chen
- Edward SunESEdward Sun
- Yihan HuangYHYihan Huang
- Shijia ZhouSZShijia Zhou
Preprint
- 4 June 2026
- preprint
- Published by Elsevier in SSRN Electronic Journal
Abstract
Roadside detector failures are often treated as missing values or noisy measurements. This paper argues that degraded sensing also creates an information-structure problem: traffic-flow predictors can lose target-specific observability of route-, horizon-, and regime-dependent latent traffic directions. We develop a sparse multimodal observability framework for traffic-flow forecasting under degraded roadside sensing. Numerical detector measurements, structured contextual variables, and text/event-derived signals are represented as separate information blocks, while vehicle-detector degradation is modeled as sparsification and sparse corruption of the numerical block. The framework defines a target-specific observability operator, posterior target uncertainty, and the critical numerical sensing density required to achieve a prescribed reliability level. The theory shows that auxiliary information lowers the critical numerical sensing density when it covers directions that become weakly observed after numerical degradation. It also predicts larger modality value under stronger detector degradation, longer forecasting horizons, and more volatile traffic regimes. The empirical study uses freeway route-level traffic-flow data to compare numerical-only, context-enhanced, text-enhanced, and context-anchored multimodal predictors. A controlled vehicle-detector degradation simulation tests whether enriched information sets maintain the same residual-risk tolerance with fewer working detectors, while the forecasting study evaluates mean error, upper-tail residual risk, horizon effects, and degradation--regime amplification. The results support the observability interpretation. Structured context consistently improves robustness relative to numerical-only sensing and increases the tolerated vehicle-detector failure rate from 0% to 30% under the clean numerical-only benchmark. Context-anchored multimodality maintains the same benchmark through the largest tested failure level of 50\%. Upper-tail gains are strongest when detector degradation coincides with high-flow, demand-surge, or large-change conditions. Text/event information is state-dependent: it can improve calibrated reliability when aligned with structured context, but it can also generate negative transfer when used without sufficient contextual anchoring. The contribution is methodological: degraded traffic sensing is formulated as sparse loss of target observability, and contextual information is characterized as an observability-restoring layer for reliable traffic-flow forecasting.Keywords
This publication has 28 references indexed in Scilit:
- Revisiting the traffic flow observability problem: A matrix-based model for traffic networks with or without centroid nodesTransportation Research Part B: Methodological, 2024
- A survey of contextual optimization methods for decision-making under uncertaintyEuropean Journal of Operational Research, 2024
- Hierarchical control for stochastic network traffic with reinforcement learningTransportation Research Part B: Methodological, 2022
- A network sensor location problem for link flow observability and estimationEuropean Journal of Operational Research, 2022
- Optimization of traffic sensor location for complete link flow observability in traffic network considering sensor failureTransportation Research Part B: Methodological, 2019
- Data-driven distributionally robust optimization approach for reliable travel-time-information-gain-oriented traffic sensor location modelTransportation Research Part B: Methodological, 2018
- Deep learning for short-term traffic flow predictionTransportation Research Part C: Emerging Technologies, 2017
- Short-term traffic forecasting: Where we are and where we’re goingTransportation Research Part C: Emerging Technologies, 2014
- Vehicular network sensor placement optimization under uncertaintyTransportation Research Part C: Emerging Technologies, 2013
- User-Friendly Tail Bounds for Sums of Random MatricesFoundations of Computational Mathematics, 2011