Modeling Non-Euclidean Relationships
During my AI research internship at the Okinawa Institute of Science and Technology (OIST) in Japan, our core objective was modeling complex ecological trophic networks. In an ecological food web, interactions are not linear sequences; they form directed, heterogeneous graphs where energy flows, predator-prey dependencies, and topological trophic nodes evolve dynamically.
Traditional convolutional and tabular machine learning methods stumble when processing graph structures because permutation invariance and variable neighborhood degree distributions break standard matrix assumptions.
GraphSAGE vs. Graph Attention Networks (GAT)
- GraphSAGE (Sample and Aggregate): Enabled inductive representation learning across large graphs by sampling uniform local neighborhoods and aggregating neighbor node features rather than processing the entire adjacency matrix.
- Graph Attention Networks (GAT): Applied multi-head self-attention over node edges, allowing the model to weight the ecological significance of specific trophic links differentially.
Multi-LLM Synthesis for Literature Bottlenecks
To complement graph topological modeling, we engineered data synthesis pipelines leveraging multi-LLM ensembles (Gemini, DeepSeek, Anthropic, and OpenAI). The system extracted unstructured interaction data from thousands of ecological research publications, transforming fragmented citations into structured edge-list datasets for graph training.