Graph Neural Network for User Personality Prediction

Developed a graph representation learning pipeline to predict personality categories from large‑scale user–product interaction data.

• Graph Construction: Projected bipartite interactions + derived structural features (degree, clustering proxies).
• Model: Stacked GraphSAGE / GAT hybrid variants evaluated; final architecture balanced expressivity & overfitting resistance.
• Feature Fusion: Combined user metadata embeddings + learned product vectors + topological encodings.
• Training: Mini-batch neighbor sampling, class imbalance reweighting, early stopping via macro-F1.
• Evaluation: Achieved Weighted F1 0.91 (↑15% vs strongest baseline MLP); robust generalization to cold nodes.

Results: Demonstrated effectiveness of relational inductive biases for sparse behavioral inference tasks.

Tech: Python, PyTorch, PyTorch Geometric.

View on GitHub →

Inference pipeline batches node embeddings and caches intermediate neighbor h-hop expansions to reduce latency.