📰 Personalized News Recommendation System Using Graph Neural Networks

📌 Overview

This project builds a personalized news recommendation system using Graph Neural Networks (GNNs) to deliver tailored article suggestions by modeling the relationships between users, articles, and content metadata. By leveraging graph-based learning, the system captures complex interactions and semantic patterns beyond traditional collaborative or content-based methods.

🧠 Problem Definition

• Goal: Recommend news articles aligned with individual user preferences.
• Challenge: Address dynamic user interests, content freshness, and the cold-start problem.
• Solution: Use GNNs to learn from a heterogeneous graph of users, articles, categories, and interactions.

📂 Dataset and Graph Construction

🔹 Datasets Used:

• News Article Data: Includes ID, title, category, sub-category, and abstract.
• User Interaction Data: Tracks which articles were shown and clicked by users.

🔹 Node Types:

• User IDs: Active and cold-start users.
• News IDs: Individual articles.
• Categories/Sub-Categories: Classification metadata.

🔹 Feature Embeddings:

• News Articles: Title + abstract → Sentence Transformer embeddings (all-MiniLM-L6-v2).
• Categories/Sub-Categories: Based on label names.
• Users:
  • Active: Average embeddings of clicked articles.
  • Cold-start: Average embeddings of popular articles.

🔹 Node Features:

• Node ID, Node Type, Node Label, Node Feature Embedding

🔹 Edge Types:

• Category ↔ Sub-category
• Sub-category ↔ Article
• User ↔ Article (click interaction)

The resulting graph represents a multi-level, heterogeneous structure of user-content relationships for GNN-based learning.

⚙️ Model Selection

🔍 GNN Architectures Explored:

1. Graph Convolutional Network (GCN): Captures global graph structures via neighborhood aggregation.
2. Graph Attention Network (GAT): Learns importance of neighbor nodes using attention weights.
3. GraphSAGE: Scalable sampling-based aggregation, ideal for inductive learning and cold-start scenarios.

🧪 Training Setup:

• 4-layer GNNs
• Link prediction task with positive (real) and negative (random) edges.
• Loss Function: Binary cross-entropy
• Epochs: 100
• Evaluation: On separate validation sets using both classification and ranking metrics.

📈 Evaluation Metrics

📊 Classification Metrics:

• Average Precision (AP)
• Average Recall (AR)
• Average F1 Score

Model Type	Avg Precision	Avg Recall	F1 Score
Popularity-Based	0.00008	0.00109	0.00016
Content-Based (Article)	0.00018	0.00855	0.00034
Content-Based (User)	0.00014	0.00098	0.00023
GCN	0.048	0.79	0.091
GAT	0.048	0.651	0.090
GraphSAGE	0.041	0.485	0.075

📌 GCN achieves the best classification performance with a strong recall and balanced F1 score.

📊 Ranking Metrics:

• NDCG@k (Normalized Discounted Cumulative Gain)
• MRR (Mean Reciprocal Rank)

Model	NDCG@5	NDCG@10	NDCG@15	MRR
GCN	0.627	0.674	0.707	1
GAT	0.630	0.677	0.709	1
GraphSAGE	0.624	0.672	0.705	1

All models rank highly relevant articles effectively, with GAT slightly outperforming others on NDCG.

✅ Conclusion

• Graph-based modeling significantly outperforms popularity and traditional content-based methods.
• GCN is the most effective model for this task, with high recall and ranking accuracy.
• GAT performs comparably and may offer further advantages with more complex or weighted relationships.
• GraphSAGE, while slightly behind in metrics, brings scalability and support for inductive inference, especially useful in real-time applications.

🛠️ Tech Stack

• Python, PyTorch Geometric, NetworkX, scikit-learn, HuggingFace Transformers (SentenceTransformers)
• Jupyter Notebooks for experiments and evaluation

👨‍💻 Author

Abde Manaaf Ghadiali Final Project – Machine Learning (CSCI 6364)