ModeConv: A novel convolution for distinguishing anomalous and normal Structural Behavior

Installation

Dependencies:

pip install torch --index-url https://download.pytorch.org/whl/cu117
pip install -r requirements.txt

How to use

Training and evaluating a model

python train.py

optional arguments:
  -h, --help           show this help message and exit
  --model MODEL        options: 'ModeConvFast', 'ModeConvLaplace', 'ChebConv', 'AGCRN', 'MtGNN' (default: 'ModeConvFast')
  --dataset DATASET    options: 'simulated_smart_bridge', 'luxemburg' (default: 'luxemburg')
  --epochs N           (default: 50)
  --batch-size N       (default: 256)
  --lr lr              initial learning rate for optimizer e.g.: 1e-4 | 'auto' (default: 'auto')
  --num-layer N        (default: 3)
  --decoder DECODER    options: 'linear' for linear layer only decoder, 'custom': to use ModeConv/ChebConv/etc. layers in decoder (default: 'custom')
  --hidden-dim N       (default: 8)
  --bottleneck N       (default: 2)
  --no-maha-threshold  mahalanobis threshold calculation/evaluation is very slow; disabling saves 30min-2h in val and test on luxemburg dataset (default: True)
  --seed N             (default: 3407)
  --no-cuda            (default: False)

Evaluating a model

python eval.py Path

positional arguments:
  Path                 path to pretrained weights, e.g. data/pretrained/luxemburg/ModeConvFast/GNN.statedict

optional arguments:
  -h, --help           show this help message and exit
  --model MODEL        options: 'ModeConvFast', 'ModeConvLaplace', 'ChebConv', 'AGCRN', 'MtGNN' (default: 'ModeConvFast')
  --dataset DATASET    options: 'simulated_smart_bridge', 'luxemburg' (default: 'luxemburg')
  --batch-size N       (default: 256)
  --num-layer N        (default: 3)
  --decoder DECODER    options: 'linear' for linear layer only decoder, 'custom': to use ModeConv/ChebConv/etc. layers in decoder (default: 'custom')
  --hidden-dim N       (default: 8)
  --bottleneck N       (default: 2)
  --no-maha-threshold  mahalanobis threshold calculation/evaluation is very slow; disabling saves 30min-2h in val and test on luxemburg dataset (default: True)
  --seed N             (default: 3407)
  --no-cuda            (default: False)

Datasets

The csv files for the luxemburg dataset only contain about 5% of the full data used. In the case of the Luxemburg dataset the 5% percent have even been to much to load it up in github. Therefore this link heads to the dataset: [https://drive.google.com/drive/folders/1KDAhKbqPGykxWSQzIfb0A4UDI5F5NZ9t?usp=drive_link) Please store the Luxemburg dataset in the data folder of this repository.

The csv files for the simulated_smart_bridge dataset only contain about 6% of the full data.

Results

Luxemburg

python train.py --model {Model} --dataset luxemburg

Model	AUC	F1
ModeConvFast	99.99	93.09
ModeConvLaplace	92.16	73.74
ChebConv	92.07	74.53
AGCRN	98.18	86.67
MtGNN	99.99	82.00

Simulated Smart Bridge

python train.py --model {Model} --dataset simulated_smart_bridge

Model	AUC	F1
ModeConvFast	92.23	87.93
ModeConvLaplace	92.43	88.07
ChebConv	82.15	83.89
AGCRN	92.26	87.76
MtGNN	91.19	86.78