Monotone operator theory provides a computationally efficient expressive neural network capable of learning long range dependencies while maintaining peak memory efficiency. The model is simplified to
where
The input features are given by
Best install using Dockerfile.mignn
, quick install with requirements.txt
.
The snap_amz
dataset can be found here.
The CGS dataset is provided by the CGS repository.
All other datasets are downloaded at runtime using the PyG library.
Benchmark tasks can be run using
CUDA_VISIBLE_DEVICES=<device> python3 /root/workspace/MIGNN/tasks/<task>.py
Porenet task can be run by placing the cgs_module into the top level CGS codebase.
To assure fixed point convergence several parametrizations of
Flag | Parametrization | MIGNN |
---|---|---|
cayley |
✅ | |
expm |
✅ | |
frob |
✅ | |
proj |
✅ | |
symm |
✅ | |
skew |
✅ |
Each of the fixed point solving methods can be found in agg._deq
several of these are operator splitting methods (OSM). Operator splitting methods require the residual operator. In addition several of the methods can be called with further acceleration schemes.
Class | OSM | Accelerated |
---|---|---|
DouglasRachford |
✅ | ❌ |
DouglasRachfordAnderson |
✅ | ✅ |
DouglasRachfordHalpern |
✅ | ✅ |
ForwardBackward |
✅ | ❌ |
ForwardBackwardAnderson |
✅ | ✅ |
PeacemanRachford |
✅ | ❌ |
PeacemanRachfordAnderson |
✅ | ✅ |
PowerMethod |
❌ | ❌ |
PowerMethodAnderson |
❌ | ✅ |
The operator splitting methods and their accelerated counterparts require the residual operator
If the graph is very large this direct inverse calculation will be exceptionally expensive. Alternative methods for reducing the order of the model given in agg._conv
are depicted below.
Method | Flag | Numerical Scheme | Complexity |
---|---|---|---|
Direct Inverse | direct |
||
Eigen Decomposition | eig |
||
Neumann Expansion | neumann-k |
@misc{
baker2023stable,
title={Stable, Efficient, and Flexible Monotone Operator Implicit Graph Neural Networks},
author={Justin Baker and Qingsong Wang and Bao Wang},
year={2023},
url={https://openreview.net/forum?id=IajGRJuM7D3}
}