City traffic operators increasingly receive congestion warnings for roads that no sensor watches. Physical sensors cover less than 10% of a city's road links (Xu et al.), so the remaining links get their state filled in. That filling can come from interpolation over time, from spatial inference using probe vehicles, or from both. A proactive system then raises an alarm when a predicted state crosses a threshold. Clinical early-warning research has documented what happens when such alarms are frequent and often wrong: operators learn to ignore them. We argue that sparse-sensor traffic monitoring faces a sharper version of this problem than the clinic does. Many false alarms there are produced by the interpolation itself, through spline overshoot, stale held-forward values, or sensors that drop out precisely during congestion. A single real congestion episode also spreads backward across many links and so generates a storm of near-duplicate alerts.
The proposal is to change two things about how alarms are issued. First, the unit of alarm becomes a congestion episode: a cluster of links and times connected on the road graph, tracked as it propagates, rather than an individual link crossing a threshold. Second, each alarm decision is conditioned on the provenance of the evidence behind it. Provenance here means how much of the triggering signal comes from real recent measurements and how much from interpolated values, and whether different interpolators agree. We hypothesize that, at matched lead time and matched recall, these two changes together cut alarm volume by a large factor relative to per-link thresholds with fixed lockouts. We also hypothesize that provenance adds value beyond a model's own predictive uncertainty, because uncertainty estimates are miscalibrated exactly where missingness is tied to congestion.
Core Claim
On a road network where most link states are inferred, a large share of non-actionable congestion alarms comes from two sources that a better forecaster does not remove. The first is reconstruction artifacts, which are concentrated where the evidence is old or the interpolators disagree. The second is spatial redundancy, where one physical shockwave is counted as many events. Alarm policies that see provenance and group alarms by episode should therefore beat both fixed-lockout suppression and thresholding on predictive uncertainty. The size of the gain should grow as sensor coverage shrinks and as missingness becomes more correlated with congestion. That scaling claim is falsifiable, and the experiments are built around it.
Where This Sits in Prior Work
The forecasting side supplies the method. Graph Neural Controlled Differential Equations for Traffic Forecasting turns irregular sensor histories into continuous paths with natural cubic splines. A controlled differential equation, which is a continuous-time recurrent model driven by the derivative of that path, then steers hidden states while a learned graph handles spatial structure. Its errors under 50% random missingness were "comparable to some baselines using 100% of the data." We use it as our forecaster. However, it is evaluated only as an open-loop regressor on fixed loop-detector networks with random dropout. It never issues alarms and never faces coverage below 10%. Coverage that sparse is exactly where cubic splines between distant points can overshoot.
Two papers motivate the provenance signal by showing that the right interpolator depends on the dynamics. Morelli et al. found that linear interpolation distorts heart-rate frequency content where quadratic interpolation does not. Gashi et al. found that simple forward-fill suffices where the signal changes slowly. Traffic contains both regimes: steady free flow and abrupt congestion onset. We therefore treat disagreement between forward-fill, linear, and cubic reconstructions as a direct, cheap measure of whether an alarm is being driven by data or by the choice of interpolator. The coverage literature supplies the spatial filling. HTTE uses Gaussian-process regression over learned link embeddings to estimate travel time on links with zero probe reports. Furtlehner et al. propagate sparse probe data by belief propagation. Neither work asks what alarms built on those estimates do to an operator.
The alarm-fatigue side both motivates and constrains the design. Yoon et al. combined a model of risk-score trajectories with a 15-minute lockout. This cut a 20-bed unit from about 159 to about 16 alerts per hour while still capturing 92.4% of hypotension events. That is the strongest evidence in the file that suppression works, and also the clearest example of a rigid suppression rule. On a road graph, a per-link lockout does not stop neighboring links from firing about the same shockwave, and a network-wide lockout would hide a genuinely new incident. Rajkomar et al. explain why low event prevalence makes alarm precision collapse. Krijthe et al. address this by optimizing partial AUC, the area under the ROC curve restricted to low false-positive rates, and by using decision-curve analysis. Their paper also flags that standard methods assume data are missing at random, which fails when missingness tracks crises. The sepsis threshold study showed that thresholds tailored to each department beat a uniform 80%-sensitivity cutoff. We borrow that idea in a minor role, setting thresholds by link criticality.
The planning papers bound the closed-loop part of the work. Pozanco et al. cut average waiting time in a congested Houston SUMO scenario from 582 to 121 seconds with predictive planning, but they assume full observability. SimplyfAI likewise assumes full macroscopic visibility. They define the full-information ceiling our sparse-sensor controller will be compared against.
Research Program
The system has three parts.
The first is a forecaster with provenance tracking. For each link and time step it outputs a predicted state and an occupancy or speed forecast at lead times of 15 to 60 minutes. It also outputs a provenance vector. One entry is the time since the last real observation on that link and its graph neighbors. A second is the fraction of the path's input window that was interpolated. A third is the spread among forecasts made from forward-filled, linear, and cubic-spline reconstructions of the same history. The fourth is the forecaster's own uncertainty, from an ensemble.
The second part is the episode tracker. Links whose forecast crosses their threshold are connected into one candidate episode when they are adjacent on the graph and their onset times are consistent with backward queue propagation. Consistency is judged against an upstream-speed bound estimated from the fully observed training data, not set by hand. An episode keeps its identity as it grows. A crossing that cannot be attached to a live episode starts a new one. This rule replaces the fixed lockout: suppression covers "the same physical event," not "the same 15 minutes." Secondary incidents that break the propagation bound therefore re-arm the alarm.
The third part is the alarm gate. A small classifier, trained on held-out episodes, decides whether to surface each new episode. Its inputs are episode summary features, including size, growth rate, peak forecast, and aggregated provenance. Its training target is whether a real congestion onset occurs on ground-truth sensors within the lead window. The operating threshold for each link class is set by a cost ratio between missed and false alarms. Following the sepsis study's departmental logic, that ratio is weighted by link criticality, using equilibrium volume-to-capacity ratios from the 20-city assignment dataset as the criticality signal.
We also run a missingness stress protocol. Sensors are dropped either uniformly at random or with probability that rises with true occupancy, which models buffer overflow and outages during congestion. This is the missing-not-at-random regime that Krijthe et al. warn about.
Baselines
The naive reference is per-link threshold alerting on the STG-NCDE forecast with no suppression. It measures the raw alarm storm, much like the roughly 159 alerts per hour that the single-stage model produced in Yoon et al..
The strongest existing suppression method is Yoon et al.'s design transplanted to links. Each link gets a stacked model over the trajectory of its risk score plus a 15-minute lockout. The reported result to beat is a reduction from about 159 to about 16 alerts per hour while still catching 92.4% of events, on MIMIC-III hypotension. No reported result was found on Lacuna for congestion-alarm precision or alarm burden on any traffic network. The traffic version of this baseline will therefore be measured by us, and not quoted from elsewhere.
The forecasting reference is STG-NCDE itself, with a reported MAE of 15.57 on PeMSD3. We will reproduce its forecast accuracy at full coverage before masking any sensors. This confirms that any alarm-level differences are not caused by a weaker forecaster.
The uncertainty-only gate is a strong competitor. It has the same episode tracker but alerts only when the lower bound of the ensemble's predictive interval crosses the threshold, and it ignores provenance. A second variant replaces the ensemble with an HTTE-style Gaussian-process posterior for spatial infill.
The ablations isolate each ingredient. One uses provenance without episodes, with per-link gates and lockouts. One uses episodes without provenance. One drops the criticality weighting and uses a single uniform threshold.
Fairness rules apply to all methods. Every method uses the same forecaster weights, the same masked sensor layout, and the same training, calibration, and test split by date. Each also gets the same budget for tuning its gating hyperparameters, with equal numbers of configurations searched on the calibration split. Beating the strongest baseline means one of two outcomes at matched event recall and matched median lead time of at least 15 minutes: at least 2× fewer alarms per operator-hour, or higher episode-level precision with non-overlapping bootstrap intervals. Either must hold in the missing-not-at-random condition, not only under random dropout.
Why not just…?
Why not just use the fixed lockout that already worked in the ICU? A per-patient lockout works because a patient is a single node. A congestion queue occupies many links, so per-link lockouts still fire once per link the queue reaches. This should show up as a high alerts-per-episode count for the Yoon-style baseline. Measuring that count is a direct test. If per-link lockouts already reach about one alarm per episode, the episode tracker is unnecessary, and we would report that.
Why not just threshold on calibrated predictive uncertainty? This is the most serious threat. If the forecaster's uncertainty were well calibrated on interpolated links, provenance would carry no extra information. We expect calibration to fail under occupancy-correlated dropout, because the model then sees calm-looking interpolated histories exactly when congestion is building. The decisive comparison is the uncertainty-only gate against the provenance gate, split by dropout regime. If the two tie under missing-not-at-random dropout, we drop the provenance claim and the contribution reduces to episode grouping.
Why not just place a few more sensors? Coverage costs money, but the question is how alarm quality scales with coverage. We sweep coverage from 5% to 30% of sensors retained. We will report the coverage at which naive alarming matches our method's burden at 10%. That number is the infrastructure saving, and it is the honest measure if the saving turns out small.
Why not just tune a better forecaster? The bandwidth-prediction study shows that better forecasters reduce error. The base-rate argument of Rajkomar et al. implies that small residual error still yields poor precision when congestion onsets are rare. We test this directly by holding the gate fixed and swapping in forecasters of different accuracy, which shows how much of the alarm burden a better forecaster can remove on its own.
Proposed Experiments
The open-loop study uses PeMSD4 and PeMSD8. We retain 5%, 10%, 20%, or 30% of sensors as model inputs and use every held-out sensor as ground truth. Congestion onset is defined as a sustained speed drop below a per-station threshold on held-out sensors. Missingness is applied to the retained sensors under both the random and the occupancy-correlated regimes. The metrics are episode-level recall at lead times of 15 or more minutes, median lead time, precision, alarms per operator-hour, and alarms per true episode. We also report partial AUC at low false-positive rates, following Krijthe et al..
The decisive check is an interaction effect. If the core claim holds, the advantage of provenance over uncertainty-only gating should widen as coverage falls and should be largest under correlated dropout. A flat or reversed pattern would falsify the claim.
The closed-loop study uses SUMO on the 35-junction Houston grid from Pozanco et al.. A proactive signal controller acts only on surfaced alarms and observes 10% of links. We compare average waiting time against the reported 582-second static baseline and the 121-second full-observability learning controller. The reported figures place our runs relative to the published ones. Under our sparse condition, we rerun the static and full-observability controllers ourselves on the same scenario so that every comparison shares one setup. We also measure how many controller actions each alarm policy triggers. The question is whether fewer, episode-level alarms recover most of the full-information benefit, or whether suppression costs waiting time that per-link alarming would have saved.
Finally, the criticality-weighted thresholds are checked on the San Francisco and Chicago networks from the 20-city dataset. There we report how expected delay cost and alarm burden trade off as the cost ratio changes, compared with a single uniform threshold.
References
- A unified dataset for the city-scale traffic assignment model in 20 U.S. cities
- Graph Neural Controlled Differential Equations for Traffic Forecasting
- Analysis of the Impact of Interpolation Methods of Missing RR-intervals Caused by Motion Artifacts on HRV Features Estimations
- Handling Missing Data For Sleep Monitoring Systems
- HTTE: A Hybrid Technique For Travel Time Estimation In Sparse Data Environments
- Propagation of information on undirected dependency graphs for road traffic inference
- Prediction of hypotension events with physiologic vital sign signatures in the intensive care unit
- Reply: metrics to assess machine learning models
- Development and validation of an early warning model for hospitalized COVID-19 patients: a multi-center retrospective cohort study
- Optimizing the Implementation of Clinical Predictive Models to Minimize National Costs: Sepsis Case Study
- Urban Traffic Control Assisted by AI Planning and Relational Learning
- Automated Planning for Urban Traffic Management
- Predicting Bandwidth Utilization on Network Links Using Machine Learning