🧠 Citeseer Hard Split Challenge

Welcome to the Citeseer Hard Split Challenge!
In this challenge, participants perform a node classification task on the well-known CiteSeer citation network dataset using Graph Neural Networks (GNNs) or any other graph-based approach, but test nodes are significantly different than the publicly used version.

---

📊 Dataset: CiteSeer

The CiteSeer dataset is a widely used benchmark in graph representation learning and node classification tasks.

It consists of:

• 3,327 nodes (scientific publications)
• 4,732 edges (citation links between publications)
• 3,703-dimensional binary node features, representing word occurrences in document abstracts
• 6 node classes, corresponding to different research topics

CiteSeer has been extensively used in state-of-the-art GNN papers, including GCN, GAT, GraphSAGE, and many others.

In the original setup, the dataset comes with predefined:

• training nodes
• validation nodes
• test nodes

These splits are commonly used to benchmark node classification performance.

---

🎯 Challenge Setup

In this challenge, we define a new and harder classification task.

Instead of using the original dataset splits, we introduce:

• train_mask_challenge
• val_mask_challenge
• test_mask_challenge

These masks define new train, validation, and test nodes, making the task more challenging than the original CiteSeer benchmark.

🔒 Hidden Labels

• Node features and graph structure are fully accessible
• However, the labels of challenge nodes are hidden
• Hidden labels are set to -1, meaning:

  You do not have direct access to the true labels of challenge nodes

Participants must infer labels purely from graph structure and node features.

---

🧪 Objective

The main goal is to achieve the highest classification accuracy on test_mask_challenge.

In addition:

• You may use the original CiteSeer masks for analysis
• Comparing the structural or feature properties between:
  • original test set
  • challenge test set
    can provide useful insights and hints.

🎯 Bonus objective:
Achieve a small performance gap between the challenge task and the original task, indicating strong generalization.

---

📁 How to Submit?

Train your model using the data inside the data folder. Predict labels for every node and write the predictions in the following format :

preds
3
1
2
...

Save it as a .csv file (e.g. my_submission.csv) in the submissions/ folder.
Note: .csv files in submissions/ are git-ignored, so your raw submission will not be pushed. You will submit an encrypted version instead.

From the project root, run the encryption script so it can find your CSV and the encryption key:

Linux / macOS:

cd submissions
python encrypt_submissions.py
cd ..

Windows (Command Prompt):

cd submissions
python encrypt_submissions.py
cd ..

Windows (PowerShell):

cd submissions
python encrypt_submissions.py
cd ..

This creates a .enc file next to each .csv in submissions/ (e.g. my_submission.csv.enc). Only .enc files are tracked by git; your .csv stays local. Please rename your .enc file such that it is github_name.enc

---

🏆 Evaluation

• Submit your .enc file with metadata.json under the submissions/
• Submissions are evaluated automatically using GitHub Actions
• True labels are stored securely and are never exposed
• Results are displayed on the public leaderboard

---

Good luck, and enjoy the challenge! 🧩

🏆 Leaderboard

The leaderboard is automatically updated when you submit your solution via Pull Request.

🏆 View Live Leaderboard

The leaderboard shows:

• Rank: Your position based on challange test set accuracy
• Team Name: Your Github username (without .csv)
• Challenge Acc: Challenge task test set accuracy
• Original Acc: Original task test set accuracy
• Gap: Performance gap between the challenge task and the original task