01 Project
Heterogeneous GNNs × Reinforcement Learning
Adaptive Recommendations
Recommendation system combining heterogeneous GNNs with reinforcement learning for adaptive ranking across user, item, and category nodes in a tripartite graph.
Problem
Static recommenders struggle when preferences change and when users, items, and categories carry different types of relationships.
Context
The project explored whether heterogeneous representation learning could provide a better state for adaptive ranking than a conventional interaction matrix.
Approach
I modeled users, items, and categories as a tripartite graph. A heterogeneous GNN learns relation-aware embeddings, while a reinforcement-learning agent updates ranking decisions from new interaction signals.
Architecture
The system separates graph construction, offline representation training, an online ranking policy, feedback collection, and evaluation against static baselines.
Results
The hybrid system improved ranking quality and responded more quickly to evolving preference signals in the evaluation environment.
Challenges
Balancing exploration with stable recommendations required careful reward design and offline safeguards.
Learnings
Adaptive ranking works best when the learned state represents relationships explicitly and the reward reflects long-term usefulness.
Repository
An anonymized implementation and evaluation notebook are being prepared for release.