This repository contains the implementation of Tramba, a Mamba-based model with adaptive attention designed for urban traffic speed prediction. It jointly models long-range temporal dependencies and non-local spatial correlations to capture complex traffic dynamics in dense urban networks.

---

Model Architecture

The Tramba model integrates:

• Adaptive attention to capture non-local spatial correlations across road links,
• Mamba to capture temporal dependencies,
• Gated fusion of spatial and temporal features.

---

Project Structure

├── data/
│   └── seoul_traffic_speed (sample).csv  # Sample 5-minute speed data for Gangnam links
│   └── GN_links.csv             # Metadata for Gangnam road link IDs
├── results/                     # Saved results will be stored here
├── data_loader.py              # Data preprocessing and sequence construction
├── models.py                   # Tramba model definition
├── run_tramba.ipynb            # Training and evaluation notebook

️ Data Description

• The dataset includes 5-minute interval speed records (in km/h) for road links in the Gangnam-gu area of Seoul, South Korea.
• It is publicly available via TOPIS upon request.
• This repository includes a sample dataset.
• Disclaimer: The author does not have the right to redistribute the full dataset.

seoul_traffic_speed.csv

This file contains traffic speed time-series with the following columns:

Column	Description
PRCS_YEAR	Year of observation
PRCS_MON	Month of observation
PRCS_DAY	Day of observation
PRCS_HH	Hour of observation
PRCS_MIN	Minute of observation
LINK_ID	Unique identifier of a road link (integer)
PRCS_SPD	Average speed in km/h

A datetime index is constructed from the time columns for temporal sorting and sequence construction.

GN_links.csv

This file filters the set of links used for training. It contains:

Column	Description
fid	Row index or feature ID
LINK_ID	Original identifier from the GIS network
s_link	Final link ID used in modeling and filtering speed observations

The s_link column lists link IDs that will be included in training. Only rows in seoul_traffic_speed.csv whose LINK_ID is in this list are used.

Note: To successfully run the model, the s_link values in GN_links.csv must include all LINK_IDs that appear in the seoul_traffic_speed.csv dataset.

Configuration

Inside run_tramba.ipynb, you can customize:

• SEQ_LIST: input sequence lengths (e.g., [36, 48])
• PRED_LIST: output horizon lengths (e.g., [1, 6, 12, 24, 36])
• D_MODEL, BATCH_SIZE, EPOCHS: model and training settings

---

Output

After training, the script will generate:

• Trained weights in results/
• Runtime logs and training loss/val loss history (.pkl)
• Evaluation metrics in CSV format

---