Research · 2026
LHP-GNN
A heterogeneous graph neural network that learns which meta-paths to follow per user with a PPO policy, aimed at cold-start recommendation.
Meta-path methods for recommendation usually fix the paths by hand: user → item → category → item. LHP-GNN learns them. An actor-critic policy trained with PPO picks meta-paths per user; relation-specific GNN layers embed along each path; multi-head attention fuses the results; an MLP scores user–item pairs with BPR or BCE loss.
Users with few interactions get demographic and attribute features, which is where learned paths help most. The implementation is pure PyTorch with a small HeteroGraph class. Evaluation covers Recall@K, NDCG@K, hit ratio, coverage, diversity and a Gini index for popularity bias.