Detect drift
Rank shared nodes using representation shift and include newly introduced sensors.
CONTINUAL TRAFFIC FORECASTING · EVOLVING SENSOR NETWORKS
CoMemNet
Continual memory for traffic systems that evolve over time.
01 · ABSTRACT
Forecasting where the network
does not stand still.
Traffic networks evolve as sensors are added, removed, and shift in distribution. CoMemNet selectively updates changed nodes while carrying compact temporal memory forward. Forecasting is adjacency-free; topology only optionally expands local updates.
Across three evolving PeMS benchmarks, CoMemNet delivers accurate 15-, 30-, and 60-minute forecasts with bounded updates.
Open data release. We release the processed evolving-network datasets PEMSD4(L) and PEMSD8(M), with documentation, through the Dataset link.
3real evolving
PeMS datasets
PeMS datasets
7yearly periods
per benchmark
per benchmark
60minute horizon
evaluated
evaluated
02 · NETWORK EXPLORER
Evolving sensor
network.
Interactive schematic anchored to PEMSD3(S), 2011–2017. Its yearly periods and selected-update counts follow the paper; node positions and links are illustrative rather than the released raw sensor topology.
StableDrift detectedNew sensor
PEMSD3(S)PERIOD 03 / 072013
42 selected
03 · METHOD
Selective updates.
Persistent memory.
CoMemNet tracks distributional change between online and EMA target representations, updates only the affected part of a changing graph, and retains adaptive temporal state for the next period.
Update locally
Optionally expand selected sensors to a bounded topology neighborhood.
Replay time
Reuse node-adaptive temporal memory instead of loading full historical sequences.
04 · MAIN RESULTS
The main result,
in full context.
Annual-average metrics from the revision manuscript. Select a benchmark to inspect reported methods at 15-, 30-, and 60-minute horizons.
13.5760-min MAE
PEMSD3(S)
PEMSD3(S)
| Protocol | Model | 15 min MAE / RMSE / MAPE | 30 min MAE / RMSE / MAPE | 60 min MAE / RMSE / MAPE | Running time Total / avg. (s) |
|---|
Values from the revision manuscript. Bold indicates the best value; lower errors are better.
05 · ABLATION
What carries the
adaptation forward?
Two controlled studies isolate the effects of drift-aware updates and node-adaptive temporal memory replay. Values are annual-average 60-minute metrics.
Drift sampling
New nodes and drift-sensitive shared nodes are both necessary for efficient adaptation.
| Variant | MAE | RMSE | MAPE |
|---|
TMRB-N
Compact temporal-state reuse and informed node selection improve the complete model.
| Variant | MAE | RMSE | MAPE |
|---|
06 · FURTHER ANALYSIS
Beyond a single
accuracy number.
The revision study further evaluates historical retention, scaling behavior, robustness, and the update budget.
Continual-learning controls · PEMSD3(S)
| Method | Current MAE | AIP | BWT | Forgetting |
|---|---|---|---|---|
| Full CoMemNet | 13.706 | 31.652 | −28.015 | 24.013 |
| No-replay | 14.419 | 30.324 | −24.199 | 20.742 |
| No-TMRB-N | 14.863 | 30.955 | −26.091 | 22.364 |
| CoMemNet-retrained | 13.910 | 32.073 | −28.720 | 24.617 |
Lower AIP and Forgetting are better; larger BWT is better.
Relative percentage advantage over current-period retraining increases as the evolving PEMS D4 network scales.
Budget sensitivityIncreasing the selected-node budget generally improves current-period prediction while increasing computation and historical interference.
RobustnessWith a fixed 2017 checkpoint, Full CoMemNet remains stronger than No-replay under missing-sensor and noise perturbations.
Resource accountingCompact temporal states and bounded updates avoid loading all historical raw sequences for each incremental period.
07 · EXTENDED EXPERIMENTS
Controlled evidence,
not just one table.
Additional revision experiments quantify the update budget, temporal-memory aggregation, distribution shifts, and controlled update-policy choices.
ρ budget sensitivity · PEMSD3(S)
Selected-node count and 12-step MAE across evolving periods. The operating point used by CoMemNet is highlighted.
| ρ | 2012 | 2014 | 2015 | 2017 |
|---|---|---|---|---|
| 0 | 182 / 12.76 | 151 / 14.07 | 65 / 16.50 | 60 / 18.78 |
| 0.03 | 246 / 12.68 | 210 / 14.05 | 92 / 14.78 | 130 / 14.89 |
| 0.05 · CoMemNet | 279 / 12.58 | 263 / 13.95 | 130 / 14.41 | 192 / 14.90 |
| 0.10 | 360 / 12.48 | 379 / 13.64 | 245 / 14.11 | 338 / 14.44 |
| 0.15 | 407 / 12.52 | 445 / 13.59 | 345 / 14.01 | 406 / 14.26 |
How to read this table
Each cell is selected nodes / 12-step MAE. A larger ρ generally selects more shared nodes and can improve current-period accuracy, but increases update cost. The highlighted ρ=0.05 is the CoMemNet operating point used in the paper.
Lower budgetefficiency ↔ accuracyHigher budget
Kmem sensitivity
Temporal-memory aggregation remains stable around Kmem=12.
Sampler and update-policy controls · PEMSD3(S)
The controlled study compares whether a bounded subset alone is sufficient (Random), whether recent error or recency explains selection, alternative feature distances (L2 / KL / JS / MMD), target branches and momenta, and selected-only versus 1/2/3-hop update neighborhoods.
