DyGFM

DyGFM (Dynamic Graph Foundation Model) is a three-stage pipeline for temporal graph learning: static GCN pre-training → dynamic TGAT pre-training (two phases) → downstream fine-tuning with MoE routing and conditional prompts.

Supported downstream datasets: genre, mooc, reddit, wikipedia.

Cross-domain rule: when you set --dataset X, pre-training uses the other three domains; fine-tuning and evaluation run on target domain X.

---

Table of Contents

• Repository layout
• Environment
• Data preparation
• Quick check
• End-to-end workflow
• Fine-tuning experiments
• Outputs and checkpoints
• Module documentation

---

Repository layout

DyGFM/ # project root (all paths resolve from here via paths.py)
├── paths.py # central path constants
├── prepare_data.py # verify inputs + build static/data/*.pt
├── processed/ # node & edge feature matrices (.npy)
├── graph_structure/ # static edge_index (.npy)
├── downstream_data/ # temporal interaction CSVs
├── static/
│ ├── pretrain.py
│ ├── embedding.py # export static routing tokens
│ ├── rebuild_static_pt.py
│ ├── data/{dataset}.pt
│ └── save_model/
├── dynamic/
│ ├── pretrain.py # phase 1
│ ├── pretrain_phase2.py # phase 2
│ ├── data/features|normal_time_pt|edge_feature/
│ ├── checkpoints/
│ └── save_model/
└── fine_tune/
├── fine_tune.py # link prediction (main)
├── fine_tune_node.py # node classification
├── fine_tune_genre.py # genre-specific link prediction
├── sentenc_branch/ # static weights + tokens
└── time_branch/ # dynamic weights + tokens

---

Environment

Component	Recommended
Python	3.8+
PyTorch	1.10+ (CUDA build if using GPU)
PyTorch Geometric	2.5.x (for static/rebuild_static_pt.py and PyG Data)
Other	numpy, pandas, scipy, scikit-learn, tqdm

Example install (adjust CUDA wheel URL for your torch version):

pip install torch numpy pandas scipy scikit-learn tqdm
pip install torch-geometric==2.5.3
# If needed for static GCN:
pip install torch-scatter torch-sparse -f https://data.pyg.org/whl/torch-1.10.0+cu113.html

Optional: swanlab is imported by some training scripts; disable or install if you use experiment logging.

---

Data preparation

1. Required initial files (not generated by training)

Place these under the DyGFM repository root. {ds} ∈ {genre, mooc, reddit, wikipedia}.

Path	Description
processed/ml_{ds}_node.npy	Node features (172-dim), row i = node id i
processed/ml_{ds}.npy	Edge features for downstream loader
graph_structure/{ds}_edge_index.npy	Static graph edges [2, E] (ground-truth topology)
downstream_data/{ds}/ds_{ds}.csv	Temporal edges: u, i, ts, label, idx, ...
downstream_data/genre/ds_genre_{1..5}.csv	Genre node-class splits (optional for genre tasks)
dynamic/data/features/{ds}.pt	Dict/tensor with node features x
dynamic/data/normal_time_pt/{ds}.pt	Temporal graph bundle for TGAT
dynamic/data/edge_feature/{ds}.pt	Edge feature tensor for dynamic pre-training

Approximate total size for all four datasets: ~5 GB (dominated by processed/ml_*.npy and edge_feature/*.pt).

2. Generated locally (no external copy needed)

Path	How to create
static/data/{ds}.pt	python prepare_data.py or python static/rebuild_static_pt.py --dataset {ds}
static/data/cache/*	Created automatically during static pre-training if caching is enabled

3. Produced by pre-training (for fine-tuning)

After static / dynamic training, copy or export artifacts as follows.

Static branch (target domain T):

Artifact	Typical source	Destination
GCN checkpoint	static/save_model/T_*.pt (best run)	fine_tune/sentenc_branch/saved_model/T.pt
Static routing tokens	static/embedding.py → static/data/embeddings/T/{src}_embeddings.pt	fine_tune/sentenc_branch/embeddings/T/{src}_embeddings.pt

{src} = each of the three pre-training domains (e.g. for T=wikipedia: genre, mooc, reddit).

Dynamic branch (target domain T):

Artifact	Typical source	Destination
TGAT checkpoint	dynamic/checkpoints/overall_best_model_T.pt (phase 1) or overall_best_model_T_phase2.pt (phase 2)	fine_tune/time_branch/saved_model/T.pt
Dynamic routing tokens	Export / training pipeline (shape [N, 64] per source domain)	fine_tune/time_branch/embeddings/T/{src}_embeddings.pt

Fine-tuning will fail if saved_model/T.pt or cross-domain *_embeddings.pt files are missing.

---

Quick check

From the repository root:

cd /path/to/DyGFM
python prepare_data.py

Expected: all [OK] lines and Saved static/data/{ds}.pt for each dataset (or [OK] if already present).

---

End-to-end workflow

Replace T with your target domain (e.g. wikipedia). Use smoke-test epoch counts first, then scale to paper settings.

Step 0 — Verify data

cd /path/to/DyGFM
python prepare_data.py

Step 1 — Static pre-training

Pre-trains on the three domains other thanT.

cd static
python pretrain.py --dataset T --gpu 0 --nb_epochs 5000 --patience 200 --lr 1e-5

Smoke test:

python pretrain.py --dataset wikipedia --gpu 0 --nb_epochs 2 --patience 50

Copy the best checkpoint:

cp save_model/wikipedia_<timestamp>.pt \
../fine_tune/sentenc_branch/saved_model/wikipedia.pt

Generate static routing tokens (run from static/, model path = your checkpoint):

python embedding.py --dataset wikipedia --model_path save_model/wikipedia_<timestamp>.pt --gpu 0
mkdir -p ../fine_tune/sentenc_branch/embeddings/wikipedia
cp data/embeddings/wikipedia/*_embeddings.pt \
../fine_tune/sentenc_branch/embeddings/wikipedia/

Step 2 — Dynamic pre-training (phase 1 + phase 2)

Phase 1 — train shared backbone, adapters frozen:

cd ../dynamic
python pretrain.py --dataset T --gpu 0 --freeze_adapter --epochs_per_domain 30 --alternating_cycles 2

Checkpoint: dynamic/checkpoints/overall_best_model_T.pt

Smoke test:

python pretrain.py --dataset wikipedia --gpu 0 --freeze_adapter --epochs_per_domain 1 --alternating_cycles 1

Phase 2 — freeze backbone, fine-tune adapters:

python pretrain_phase2.py --dataset T --gpu 0 \
--phase1_model_path checkpoints/overall_best_model_T.pt \
--epochs_per_domain 50

Copy TGAT weights for fine-tuning:

cp checkpoints/overall_best_model_T_phase2.pt \
../fine_tune/time_branch/saved_model/T.pt
# Or use phase-1 weights if phase 2 is skipped:# cp checkpoints/overall_best_model_T.pt ../fine_tune/time_branch/saved_model/T.pt

Place dynamic routing tokens under fine_tune/time_branch/embeddings/T/ (one {src}_embeddings.pt per source domain). These are produced by your dynamic export / pre-training pipeline.

Step 3 — Fine-tuning

cd ../fine_tune
python fine_tune.py --dataset T --gpu 0 --epochs 1000 --num_runs 100 --patience 100 --val_freq 5

Smoke test (requires weights + tokens from steps 1–2):

python fine_tune.py --dataset wikipedia --gpu 0 --epochs 1 --num_runs 1 --val_freq 1

---

Fine-tuning experiments

All commands assume cd fine_tune and that sentenc_branch/ + time_branch/ artifacts exist for the target domain.

Link prediction (main)

python fine_tune.py \
--dataset wikipedia \
--gpu 0 \
--epochs 10000 \
--num_runs 100 \
--patience 100 \
--val_freq 5 \
--prefix exp_wiki

Metrics: transductive / inductive AP and AUC on validation and test splits (see console and fine_tune/results/).

Node classification

Uses ds_{genre}_{class}.csv when --dataset genre.

python fine_tune_node.py \
--dataset genre \
--genre_class 1 \
--gpu 0 \
--epochs 10000 \
--num_runs 100 \
--patience 100 \
--train_shot_num 3 \
--val_shot_num 3

Smoke test:

python fine_tune_node.py --dataset genre --genre_class 1 --gpu 0 --epochs 5 --num_runs 1 --patience 1000

Genre link prediction (stratified sampling)

python fine_tune_genre.py --dataset genre --gpu 0 --epochs 300 --num_runs 1

Reload / time-slice node test

python reload_test_node.py --dataset mooc --gpu 0 --task_num 5

---

Outputs and checkpoints

Stage	Location	Naming
Static pre-train	static/save_model/	{target}_{timestamp}.pt
Dynamic phase 1	dynamic/checkpoints/	overall_best_model_{target}.pt
Dynamic phase 2	dynamic/checkpoints/	overall_best_model_{target}_phase2.pt
Dynamic phase 1/2 saves	dynamic/save_model/	tgat_{target}_{timestamp}_phase{1,2}.pt
Fine-tune run logs	fine_tune/results/	{dataset}_{prefix}_run_*.pkl, *_final_summary.pkl
Fine-tune checkpoints	`fine_tune/checkpoints/edge	node/`

Path resolution is centralized in paths.py; scripts do not rely on ../ relative paths.

---

Module documentation

• Static pre-training
• Dynamic pre-training
• Fine-tuning