nglaze00/SCoNe_GCN

Using Hodge graph neural networks for path prediction; ICML paper @ https://arxiv.org/abs/2102.10058

Simplicial Complex Net (SCoNe)

A convolutional neural network architecture for trajectory prediction on simplicial complexes (i.e. graphs). In addition to node- and edge-level data, uses higher-order structures (triangles) in the graph to train very generalizable trajectory prediction models. Paper can be found on arXiv.

Use

1. Clone this repo
   • Dependencies: Python 3.7; numpy, matplotlib, scipy, networkx, jax, jaxlib, treelib)
2. Set up a dataset in one of two ways (see synthetic_data_gen.py for more info):
   • Generate a synthetic dataset (graph + trajectories) using synthetic_data_gen.py
   • Convert your own data to the format SCoNe accepts
3. Train a model (see trajectory_experiments.py for more info)
   • SCoNe: Run trajectory_experiments.py (with arguments)
   • Projection: Run projection_model.py (see file for more info)
   • Markov: Run trajectory_experiments.py with arg -markov 1
   • SNN (Ebli 2010): Run trajectory_experiments.py with arg -model 'ebli'
   • SCCONV (Bunch 2012): Run trajectory_experiments.py with arg -model 'bunch'