LARS-research/RED-GNN — reverse-engineered prompt

Build me a clean research codebase for RED GNN, a neural knowledge graph reasoning project based on relational directed graphs. I want to be able to train and test it on static knowledge graphs in both transductive and inductive settings, and also on temporal knowledge graphs for interpolation and extrapolation.

Please make it easy to reproduce experiments from the command line. Include runnable training scripts for datasets like YAGO, FB237, ICEWS14, ICEWS05 15, Wikidata11k, and YAGO temporal forecasting if the data is available. The static setup should support splitting training triples into facts and train files so queries do not leak into facts, using a simple ratio like 3 to 1.

Use PyTorch and torch scatter as the main dependencies, and add clear setup and run instructions in the README. Keep the project organized so someone can go into the static or temporal section and run the right script without digging around.