Dynamic Link Prediction with DyGLib

Overview

This project uses Dynamic Graph Library (DyGLib) to study dynamic link prediction on temporal graphs. We focus on the DyGFormer backbone and two variants:

• NWI (Non-Uniform, data-aware Weight Initialization) time encoder
• Gated SwiGLU feed-forward network

All training and evaluation code is adapted from DyGLib.

---

1. Environment Setup

We use Conda with an environment.yml file.

From the repository root:

conda env create -f environment.yml
conda activate dyglib-env

After activation, you should be able to run:

python -c "import torch; print(torch.__version__)"

to verify that PyTorch is correctly installed.

---

2. Datasets and Preprocessing

We use the datasets from:

"Towards Better Evaluation for Dynamic Link Prediction" https://openreview.net/forum?id=1GVpwr2Tfdg

All datasets can be downloaded from the Zenodo record:

https://zenodo.org/record/7213796

Place them under the DG_data folder in the repository root, e.g.

DG_data/
  wikipedia/
  reddit/
  mooc/
  uci/
  CanParl/
  USLegis/
  ...

2.1 Standard preprocessing

From the repository root:

cd preprocess_data/

preprocess a single dataset

python preprocess_data.py --dataset_name wikipedia

or preprocess all datasets

python preprocess_all_data.py

2.2 Dataset statistics for NWI

The NWI time encoder uses simple dataset-level statistics.

cd preprocess_data/
python data_statistics.py

Use the printed statistics to configure the NWI time encoder in models/modules.py. After this, you can use --time_encoder nwi during training.

---

3. Dynamic Link Prediction: Running the Code

All experiments are dynamic link prediction tasks using DyGLib’s pipeline.

3.1 Single run example

From the repository root:

python train_link_prediction.py \
  --dataset_name wikipedia \
  --model_name DyGFormer \
  --patch_size 2 \
  --max_input_sequence_length 64 \
  --num_runs 5 \
  --act_fn gelu \
  --time_encoder original \
  --gpu 0

Important flags:

• --dataset_name: wikipedia, reddit, mooc, uci, canparl, ...
• --model_name: DyGFormer, GraphMixer ...
• --act_fn: gelu or swiglu
• --time_encoder: original or nwi
• --gpu: GPU index

To run our variants:

• NWI time encoder: --time_encoder nwi
• SwiGLU FFN: --act_fn swiglu

---

4. Weights & Biases (wandb)

Training runs are logged to Weights & Biases.

export WANDB_API_KEY=YOUR_API_KEY_HERE

Wandb will record losses and metrics (e.g., validation/test AP) for each run.

---

5. Batch Experiments

We provide a simple script to compare standard configurations across datasets:

batch_run.sh

The default configuration list inside the script is:

configs=(
  "gelu original"   # baseline: GeLU + original time encoder
  "gelu nwi"        # NWI time encoder
  "swiglu original" # SwiGLU FFN
)

Edit batch_run.sh to change datasets or GPU IDs, then run:

bash batch_run.sh

---

6. Results and Analysis

All raw metrics are stored under:

saved_results/
  DyGFormer/
    ...

To aggregate metrics into CSV tables: python aggregate_metrics.py

To inspect which metrics are logged in wandb: python inspect_wandb_run_columns.py

To plot validation AP convergence curves: python wandb_api_plot.py, the generated plots will be saved to wandb_analysis/