The following figure shows the trend of loss decrease and AUC increase during model training:

https://github.com/inneedloveBu/wikinet-link-prediction/animations/training_progress_english_20260124_160006.gif

WikiNet is a graph representation learning project designed to perform link prediction on a Wikipedia hyperlink network.

The task is formulated as a binary classification problem over node pairs.

The project focuses on:

• Structural feature engineering
• Hardness-aware negative sampling
• Multi-interaction edge decoding
• Robust training stabilization techniques

The graph is constructed from the WikiLinks dataset.

• Number of nodes: 1000
• Number of edges: 5000
• Node feature dimension: 20 (structural + content features)

• Training edges: 1771
• Validation edges: 253
• Test edges: 507

• Normalized degree
• Clustering coefficient
• Betweenness centrality
• PageRank score
• Neighbor statistics (mean / std / max / min degree)

• Title length
• Word count
• Numeric token count
• Uppercase ratio
• Special character ratio
• Hash-based embedding component

All features are standardized before training.

A 3-layer MLP encoder:

• Linear → BatchNorm → ReLU → Dropout

• Linear → BatchNorm → ReLU → Dropout

• Linear projection layer

• Hidden dimension: 64

• Total trainable parameters: 21,729

Instead of naive concatenation, the decoder integrates multiple interaction patterns:

• Node embedding u
• Node embedding v
• Absolute difference |u − v|
• Element-wise product u ⊙ v

These representations are concatenated and passed through an MLP classifier.

This design increases expressive power for link prediction tasks.

Three difficulty levels were implemented:

• Easy — Random negative sampling
• Medium — Common-neighbor-based sampling
• Hard — Degree-similarity-based sampling

Final training uses:

• Medium hardness sampling

This generates more informative negative edges and improves generalization.

• Binary Cross Entropy with logits

• Class-weight balancing

• Gradient clipping

• Early stopping (patience = 20)

• ReduceLROnPlateau learning rate scheduler

• Best model checkpointing

• Training epochs: 300

• Best validation AUC: 0.897 (epoch 110)

• ROC-AUC: 0.798
• Average Precision (AP): 0.784
• F1-score: 0.763
• Accuracy: 0.696

The model demonstrates stable generalization performance under controlled negative sampling.

• PyTorch 2.0+
• PyTorch Geometric
• NetworkX
• NumPy
• Pandas
• scikit-learn
• Matplotlib

wikinet-link-prediction/
├── data/
│   ├── raw/
│   └── cleaned/
├── models/
├── notebooks/
├── train11.py
├── requirements.txt
└── README.md

git clone https://github.com/inneedloveBu/wikinet-link-prediction.git
cd wikinet-link-prediction

pip install -r requirements.txt

python train11.py

• Designed hardness-controlled negative sampling mechanism
• Implemented multi-interaction edge decoder
• Combined structural and lightweight content features
• Applied full training stabilization pipeline
• Conducted systematic multi-metric evaluation

• Replace MLP encoder with GCN / GAT
• Neighbor sampling for scalability
• Larger-scale graph experiments
• Embedding visualization
• Interactive web deployment

• MIT License

• GitHub Issues: https://github.com/inneedloveBu/wikinet-link-prediction/issues
• Email: indeedlove@foxmail.com