A Python Library for Graph Outlier Detection (Anomaly Detection)

Hi,

Given a graph object in the prediction API, What does the outlier labels mentioned hereas outlier_labels (numpy array of shape (n_samples,)) indicate from a graph perspective?

Does the contents in the numpy array as 1 or 0 indicate the Nodes in the graph that are normal or anomalous? for example Labels: [0 0 0 ... 0 0 0] . Does each 0 value pertain to a node in graph?

So, How should this prediction output be interpreted from a graph perspective? Thanks in advance.

When I use 1080ti GPU, there is an out of memory problem in all datasets except Cora, which is inconsistent with the description in the benchmark paper.

Describe the bug

The current dependencies include installing an external library: argparse.

It must be noted that argparse is a part of the core python libraries. There is no need for installing it alike other external libraries.

• docs for core-library, argparse: https://docs.python.org/3/library/argparse.html

:fire: EDIT: The library (argparse) that you are installing from PyPI is no longer maintained as it is now a part of standard python3. See my comment here.

See further details:

• https://gitter.im/conda-forge/conda-forge.github.io?at=62598b5b0466b352a46afd25

https://github.com/pygod-team/pygod/blob/d037b67bd3001f4d45be5093b3717700aa79d953/requirements.txt#L1-L5

Running examples\adone.py for replication

C:\Users\yuezh\Anaconda3\envs\torch19\python.exe C:/Users/yuezh/PycharmProjects/pygod/examples/adone.py training... Traceback (most recent call last): File "C:/Users/yuezh/PycharmProjects/pygod/examples/adone.py", line 35, in model.fit(data) File "C:\Users\yuezh\PycharmProjects\pygod\pygod\models\adone.py", line 158, in fit act=self.act).to(self.device) File "C:\Users\yuezh\PycharmProjects\pygod\pygod\models\adone.py", line 331, in init act=act) TypeError: init() got an unexpected keyword argument 'in_channels'

When I am reading paper and PyGOD code, I find a problem when some algorithms aim to reconstruct structural infomation:

$$ \hat{A}=\sigma(\pmb z\pmb z^T) $$

where z is the graph embedding we have learnt, $\sigma$ is sigmoid function, and $\hat{A}$ is reconstructed adjacency matrix. One term of objective function is

$$ \Vert A-\hat{A}\Vert_F^2 $$

where $A$ is the adjacency matrix. But we should find that the diagonal elements of $\hat{A}$ is closed to 1 because

$$ \hat{A}_{ii}=\sigma(z_iz_i^T) $$

So I think we should add a self-loop on $A$ when reconstruction:

$$ \Vert(A+I)-\hat{A}\Vert_F^2 $$

In PyGOD code, I haven't found this consideration. I modified the code of DOMINANT in this way, and found performance improvement in some dataset.

Describe the bug Hi, I am trying to run example of PyGOD in a subprocess and it does not work for me

To Reproduce

from torch.multiprocessing import Process

import torch_geometric.transforms as T
from torch_geometric.datasets import Planetoid

import torch
from pygod.generator import gen_contextual_outliers, gen_structural_outliers
from pygod.utils import load_data
from pygod.models import AnomalyDAE



def f(data):
    model = AnomalyDAE()
    print('started model fitting')
    model.fit(data)
    print('model fit succesful')

if __name__ == '__main__':
    data = Planetoid('./data/Cora', 'Cora', transform=T.NormalizeFeatures())[0]
    data, ya = gen_contextual_outliers(data, n=100, k=50)
    data, ys = gen_structural_outliers(data, m=10, n=10)
    data.y = torch.logical_or(ys, ya).int()

    data = load_data('inj_cora')
    data.y = data.y.bool()
    p = Process(target=f, args=(data,))
    p.start()
    p.join()

Expected behavior

The model does not fit for me

Desktop (please complete the following information):

• OS: all os and systems
• python: 3.8

It appears that it would throw an error if the input x contains negative values. If this is expected, we should probably mention it somewhere.

check pygod/test/test_one.py

Describe the bug When calling pygod.utils.load_data(), sometimes it returns the following error message: ConnectionError: ('Connection aborted.', RemoteDisconnected('Remote end closed connection without response'))

To Reproduce Steps to reproduce the behavior:

1. Go to '...'
2. Click on '....'
3. Scroll down to '....'
4. See error

Expected behavior A clear and concise description of what you expected to happen.

Screenshots If applicable, add screenshots to help explain your problem.

Desktop (please complete the following information):

• OS: [e.g. iOS]
• Browser [e.g. chrome, safari]
• Version [e.g. 22]

Smartphone (please complete the following information):

• Device: [e.g. iPhone6]
• OS: [e.g. iOS8.1]
• Browser [e.g. stock browser, safari]
• Version [e.g. 22]

Additional context Please refer to 1 and 2 for potential fixing approaches.

• add edge drop probability to structural outlier injection.
• update benchmark script with more datasets.
• multiple minor fixes by @cshjin @YingtongDou @kayzliu

Is your feature request related to a problem? Please describe. For now, some detectors (e.g., GUIDE) has two separate autoencoders for attribute and structure, but two autoencoders share the same hidden layer dimension. In many cases, there are a significant difference between the dimension of the node attributes and the dimension of structure information (e.g., adjacency matrix). Using the same hidden dimension may hampers the performance of the detectors.

Describe the solution you'd like Enabling different hidden dimension for attribute autoencoder and structure autoencoder

We can add a tutorial in the document about loading data from numpy, scipy, matlab, networkx and other common data formats. Some of the data loaders can be found in https://pytorch-geometric.readthedocs.io/en/latest/modules/utils.html

Hi, Thank you for this awesome package. I am working with heterogeneous and knowledge graphs. For example, if I use the famous MovieLens dataset and construct a heterogeneous graph, Can I feed it to model.fit(data) ?

When running model.fit (on GPU), I received the following error: ValueError: Expected more than 1 value per channel when training, got input size torch.Size([1, 64])

Any suggestion to get rid of this problem when using on GPU is appreciated.

The Flickr dataset used in some god papers is the Flickr in pyg. And inj_ Flickr implemented in your library has a very different dataset. I hope you can use a correct data set.

The codes for masking the target nodes is wrong. The target node is the first node in subgraph after the RandomWalk sample, while you mask the last node. The performance of CoLa and ANEMONE will improve 2% by fixing the bug.

Wrong codes in CoLA(line 361~364) batch_feature = torch.cat( (batch_feature[:, :-1, :], added_feat_zero_row, batch_feature[:, -1:, :]), dim=1)

Correct codes: batch_feature = torch.cat( (added_feat_zero_row, batch_feature[:, 1:, :], batch_feature[:, 0:1, :]), dim=1)

Wrong codes in ANEMONE(line 288~289 and 429~430) bf = torch.cat( (bf[:, :-1, :], added_feat_zero_row, bf[:, -1:, :]), dim=1)

Correct codes: bf = torch.cat( (added_feat_zero_row, bf[:, 1:, :], bf[:, 0 : 1, :]), dim=1)

Hi, could you please provide the function that returns the trained node embeddings so that I can input the embeddings to machine learning classifier such as SVM.

Best wish!