QGCN

QGCN method for graph classification: https://arxiv.org/abs/2104.06750

Installation

required packages:

• scipy~=1.8.0
• pandas~=1.4.2
• networkx~=2.8.3
• numpy~=1.22.3
• torch~=1.11.0
• scikit-learn~=1.1.1
• bokeh~=2.4.2
• matplotlib~=3.5.1
• bitstring~=3.1.9
• python-louvain~=0.16
• graph-measures~=0.1.44

You can download the package by the command:

pip install QGCN

Graph representing

To use this package you will need to provide the following files as input:

• Graphs csv file: files that contain the graphs for input and their labels. The format of the file is flexible, but it must contain headers for any column, and there must be a column provided for:
  • graph id
  • source node id
  • destination node id
  • label id (every graph id can be attached to only one label)

• External data file: external data for every node (Optional) The format of this file is also flexible, but it must contain headers for any column, and there must be a column provided for: note!! every node must get a value
  • graph id
  • node id
  • column for every external feature (if the value is not numeric then it can be handled with embeddings)

Example for such files:

• graph csv file:

g_id,src,dst,label
6678,_1,_2,i
6678,_1,_3,i
6678,_2,_4,i
6678,_3,_5,i

• External data file:

g_id,node,charge,chem,symbol,x,y
6678,_1,0,1,C,4.5981,-0.25
6678,_2,0,1,C,5.4641,0.25
6678,_3,0,1,C,3.7321,0.25
6678,_4,0,1,C,6.3301,-0.25

Parameters passing

After creating these file, you should define the parameters of the model. This can be done with a json file, or with data classes. The parameters split to 4 groups:

• graphs_data:

  • file_path - the path to the graph csv file (with the edges and labels for each graph)
  • graph_col - the name of the column with the graph id
  • src_col - the name of the column with the source node of the edge
  • dst_col - the name of the column with the target node of the edge
  • label_col - the name of the column with the label of the graph
  • directed - indicates if the graph is directed (gets True/False)
  • features - list of topologic features which will be calculated to the nodes.
    • The options are - ["DEG", "CENTRALITY", "BFS"]
    • You can read more about it here >>
  • adjacency_norm - the norm which will be used (get examples)
    • The options are - "NORM_REDUCED", "NORM_REDUCED_SYMMETRIC", "IDENTITY", "RAW_FORM"
  • standardization - the standardization which will be used
    • The options are - "zscore", "min_max", "scale"

• external:

  • file_path - the path to the external data csv file (with other node features)
  • graph_col - the name of the column with the graph id
  • node_col - the name of the column with the node id
  • embeddings - a list with the names of the embeddings features of the nodes
  • continuous - a list with the names of the continuous features of the nodes

• model:

  • label_type - 'binary' if the predication in binary, 'multi' else
  • num_classes - number of label types
  • use_embeddings - if the model should use the embeddings features (gets True/False)
  • embeddings_dim - a list with the dimensions of the embeddings features
  • activation - the activation function which will be used.
    • Notice that the activation function will be combined with SRSS function.
    • The options are - "relu_", "tanh_", "sigmoid_", "srss_"
  • dropout - the dropout rate of the model
  • lr - the learning rate of the model
  • optimizer - the optimizer of the model
    • The options are - "ADAM_", "SGD_"
  • L2_regularization - the L2_regularization rate of the model
  • GCN_layers - an array with dictionaries for each layer.
    • for example: [
      { "in_dim": "None", "out_dim": 100 },
      { "in_dim": 100, "out_dim": 50 },
      { "in_dim": 50, "out_dim": 25 }
      ]

• activator:

  • epochs - the epochs number of the model
  • batch_size - the size of each batch
  • loss_func - the loss function which will be used
  • train - percentage of the data which will used for train
  • dev - percentage of the data which will used for dev
  • test - percentage of the data which will used for test

• Example json file:
  • (Notice that if an external file is not provided, you should put the associated parameters as None.)
  • you can find complete params files here.

{
    "dataset_name": "DataSetName",

    "external": {
       -- external params here -- 
    },

    "graphs_data": {
        -- graphs_data here --
    },

    "model": {
        -- model params here --
    },

    "activator": {
        -- activator params here -- 
    }
}

• Example dataclass objects:
  • The dataclasses default values are here.

from QGCN.params import GraphsDataParams, ExternalParams, ModelParams, ActivatorParams 

external_params = ExternalParams(file_path="./data/Mutagenicity_external_data_all.csv",
                          embeddings=["chem"],
                          continuous=[])

graphs_data_params = GraphsDataParams(file_path="../src/QGCN/data/Mutagenicity_all.csv",
                               standardization="min_max")

model_params = ModelParams(label_type="binary",
                    use_embeddings="True",
                    embeddings_dim=[10],
                    activation="srss_",
                    GCN_layers=[
                        {"in_dim": "None", "out_dim": 250},
                        {"in_dim": 250, "out_dim": 100}])

activator_params = ActivatorParams(epochs=100)

Executing the model

Once you have these files, you can use the QGCNModel from QGCN.activator with the path to the parameters file or the dataclass objects:

from QGCN.activator import QGCNModel, QGCNDataSet

qgcn_model = QGCNModel(dataset_name="Aids", params_file="params.json")
qgcn_model.train()

from torch.utils.data import DataLoader
from QGCN.params import GraphsDataParams, ExternalParams, ModelParams, ActivatorParams 
from QGCN.activator import QGCNModel, QGCNDataSet

# sets the parameters of the dataset:
graphs_data = GraphsDataParams(file_path="./data/data_all.csv",
                               standardization="min_max")
external = ExternalParams(file_path="./data/external_data_all.csv",
                          graph_col="g_id", node_col="node",
                          embeddings=["chem"], continuous=[])


# sets the parameters of the model:
model = ModelParams(label_type="binary", num_classes=2, use_embeddings="True", embeddings_dim=[10],
                    activation="srss_", dropout=0.2, lr=0.005, optimizer="ADAM_", L2_regularization=0.005, f="x1_x0")
activator = ActivatorParams(epochs=100)

qgcn_model = QGCNModel("Mutagen", graphs_data, external, model, activator)
qgcn_model.train(should_print=True)

Links

The datasets can be download here: https://ls11-www.cs.tu-dortmund.de/staff/morris/graphkerneldatasets . Notice you will have to change their format to ours. You can see an example data here (gitHub link) the conventor in datasets -> change_data_format.py Mail address for more information: 123shovalf@gmail.com