CAMEL: Cross-temporal Anchor-guided Memory with Evolving Latent Dynamics

Official implementation for ultra-long-span traffic forecasting with temporal gaps.

1. Overview

Traffic forecasting under multi-year observation gaps is fundamentally different from standard short-horizon forecasting. CAMEL is designed for this setting and integrates three complementary components:

• Cross-Year Episodic Memory (CEM): retrieves season-consistent cross-year patterns as anchors.
• Latent Dynamics Extrapolator (LDE): propagates latent states through unobserved gaps via time-conditioned Graph-ODE dynamics.
• Anchor-Temporal Fusion (ATF): adaptively fuses memory, latent dynamics, and recent observations with gap-aware gating.

The method targets the gap forecasting protocol on long-span traffic benchmarks (e.g., XXLTraffic subsets such as PeMS03 / PeMS08 / TfNSW).

2. Repository Structure

CAMEL/
├── run.py                          # training/testing entry
├── run.sh                          # batch script (switch dataset/gap via DATA_FILE / GAP_DAYS)
├── exp/
│   ├── exp_basic.py               # experiment base
│   └── exp_long_term_forecasting.py
├── logs/                           # training logs used in the paper (main table, ablation, strategies, etc.)
├── models/
│   ├── CAMEL.py                   # CAMEL core model
│   ├── Autoformer.py
│   ├── DLinear.py
│   ├── FEDformer.py
│   ├── Informer.py
│   ├── PatchTST.py
│   ├── iTransformer.py
│   ├── PhaseFormer.py
│   ├── MixLinear.py
│   ├── FreqCycle.py
│   ├── stgcn.py
│   ├── astgcn.py
│   ├── gwn.py
│   └── pdformer.py
├── data_provider/
│   ├── data_factory.py
│   ├── data_loader.py
│   ├── uea.py
│   └── m4.py
├── layers/
└── utils/

3. Training

Batch run

bash run.sh

Switch dataset or gap range

run.sh supports quick switching of dataset file and gap span:

# example: run on PeMS03 with 1.5y and 2y gaps
DATA_FILE=pems03_all_common_flow.csv GAP_DAYS="548 730" bash run.sh CAMEL

You can also set:

• DATA_ROOT to change dataset directory.
• DATASET_TAG to override experiment name prefix in model_id.

4. Supported Baselines in This Repo

• MLP-based: DLinear, FreqCycle, MixLinear
• Transformer-based: Informer, Autoformer, FEDformer, PatchTST, iTransformer, PhaseFormer
• Graph-based: stgcn, astgcn, gwn, pdformer
• Proposed method: CAMEL

5. Datasets

• XXLTraffic Benchmark: https://1drv.ms/f/s!Au2cJRs-_u93hvMeZJBXJROA3lrHlA?e=3JJlt5