Replication package explanatory file

Journal title: Communications in Transportation Research (COMMTR)
Manuscript title: Learning from Yesterday's Error: An Efficient Online Learning Method for Traffic Demand Prediction
Manuscript ID: [to be filled after ScholarOne submission]
Authors: [to be filled in author order]

Replication package access
The complete replication package is deposited on Hugging Face Datasets (processed data under CC0; code under MIT) and mirrored on GitHub. After ETS-Data assigns a DOI, replace the placeholder below and also update the short summary in the manuscript (before Acknowledgements).

Hugging Face: https://huggingface.co/datasets/tjtrans/FORESEE
ETS-Data: https://doi.org/10.26599/ETSD.XXXX.XXXXXXX
Code mirror: https://github.com/xiannanhuang/FORESEE

Unzip FORESEE_replication_package.zip. All relative paths below are relative to the FORESEE/ root.

The archive contains:
1. dataset/<dataset_name>/train.npy, val.npy, test.npy, adj_mx.npy for nycbike, nyctaxi, chibike, chitaxi, bosbike, baybike, torbike.
2. saved_models/ checkpoints that reproduce the paper tables (14 files):
   STGCN and GWNET on all seven datasets.
3. models/*_config.yaml and model source files.
4. Source codes: train.py, foresee.py, eval_reproduce.py, fintune_opencity.py, Dataset.py, modules under models/ and tools/, and OpenCity configs in opencity/*.yaml.
5. expected_results/paper_metrics.csv for checking reproduced MAE/RMSE.

Data description
We provide processed secondary demand tensors, not raw trip records. These files are sufficient to reproduce the verified Ori and FORESEE numbers listed above.

Each dataset folder contains:
- train.npy / val.npy / test.npy: hourly inflow/outflow demand. Accepted shapes: (nodes, days, 24, 2) or hour-first (hours, nodes, 2). The last dimension is 2 (pick-up and drop-off, or equivalent inflow/outflow).
- adj_mx.npy: spatial adjacency among zones, float32, shape (N, N).

The tensors are aggregated from publicly released bike-sharing and taxi trip records of New York City, Chicago, Boston, the San Francisco Bay Area, and Toronto. We split them chronologically into train/validation/test and drop zones whose training-period mean demand is <= 2 (paper Section 5.1.1). After this filter, valid zone counts are: nycbike 74, nyctaxi 125, chibike 48, chitaxi 28, bosbike 43, baybike 36, torbike 36.

Raw trip-level microdata are not redistributed because they are large and already public from the original operators or open-data portals. The processed tensors fully determine the reported metrics.

Code description
Required environment:
- OS: Windows or Linux
- Python >= 3.8
- GPU recommended (device: cuda:0 in YAML); CPU works if device is set to cpu
- packages listed in requirements.txt: torch, numpy, scikit-learn, pyyaml, scipy, pandas, tqdm, statsmodels

Install:
    pip install -r requirements.txt

Main reproduction (no retraining):
    python foresee.py
This evaluates the 14 verified checkpoint pairs and writes online.csv. Compare with expected_results/paper_metrics.csv.

Subset evaluation:
    python eval_reproduce.py --jobs stgcn --out logs/reproduce_stgcn.jsonl --device cuda:0
    python eval_reproduce.py --jobs gwnet --out logs/reproduce_gwnet.jsonl --device cuda:0

FORESEE online settings (foresee.py, class Online_day):
- residual experts with EMA coefficients [0.7, 0.8, 0.9, 1.0]
- expert-weight temperature eta=10
- adaptive spatiotemporal smoothing: graph diffusion (alpha init 0.01) and temporal kernel [0.2, 0.6, 0.2], updated by SGD (lr=0.01)
- test loader batch size = 24 (one day); the residual is updated once per day using yesterday's error
- look-back window = 6 hours; prediction horizon = 1 hour


OpenCity backbone (optional)
The main STGCN / GWNET tables do not require OpenCity. To reproduce the OpenCity experiments we provide fintune_opencity.py and the dataset configs under opencity/*.yaml. The official OpenCity pretrained weights are not redistributed here; download OpenCity-base.pth from the OpenCity authors and place it at opencity/OpenCity-base.pth.

Step 1. Fine-tune OpenCity on a target dataset. In fintune_opencity.py, set
    config_path = r'opencity/<dataset>.yaml'
and uncomment
    torch.save(model.state_dict(), 'finetuned_opencity_<dataset>.pth')
Default example is chibike. Then run:
    python fintune_opencity.py

Step 2. Run FORESEE on the fine-tuned OpenCity model. In foresee.py, switch init_model to
    llm='opencity'
    config_path = r'opencity/<dataset>.yaml'
and load the matching checkpoint, e.g.
    torch.load(r'finetuned_opencity_<dataset>.pth')
When constructing Online_day and calling online(), also pass llm='opencity' so that the OpenCity forward signature is used. Then run:
    python foresee.py

Change <dataset> among nycbike, nyctaxi, chibike, chitaxi, bosbike, baybike, torbike. Keep device in the corresponding yaml consistent with the local GPU.

Simulation software description
Not applicable. This study does not use traffic microsimulation. All experiments are supervised forecasting on historical demand tensors.

Experiment design description
Not applicable. There are no surveys, questionnaires, or human-subject experiments. Evaluation is rolling hourly prediction on the held-out test split, reported as MAE and RMSE after inverse standardization.

Others
1. Zone filter and adjacency indexing must stay consistent: paper_valid_grid() selects zones and load_paper_adj() slices adj_mx with the same index.
2. Features are standardized with a scaler fitted only on training data (CustomStandardScaler); MAE/RMSE are computed after inverse_transform.
3. Set device in each models/*_config.yaml to match the local machine.
4. All 14 STGCN / GWNET checkpoints are released.