Neural network for subgraph classification

An embedding and CNN classification algorithm for subgraph classification.

Algorithm

• Input = subgraphs, subgraph labels, node colors

• For each subgraph:

  1. Compute graph embedding using node2vec (random walks + word2vec algorithm), ndimensions = 128
  2. Reduce to a 2D dimensional space discretized into a 2D grid using generative topographic mapping (GTM), ugtm implementation
  3. For a subgraph 2D image (grid), the first channel is node density, the other channels covariates

• Run CNN classification algorithm, with following layers (this architecture will change in the future):

  1. ZeroPadding2D((3, 3))
  2. Conv2D(32, (5, 5), strides=(1, 1))
  3. BatchNormalization
  4. Relu activation
  5. MaxPooling2D((2, 2))
  6. Flatten
  7. Dense layer with sigmoid activation

Run example (10-fold cross-validation)

python Graph2Image_CV.py

Use your own files

python Graph2Image_CV.py --input list_train_test --output output --labels random_labels --colors example_colors

Input format description

--input

List of paths to your subgraphs (one per line). The format of each subgraph should be space-separated, without header, and with 3 columns (node1 node2 weight).

--output

Just the output name.

--labels

Binary labels (0/1) for subgraphs, one per line (name number of lines as the --input file).

--colors

Covariate, with 2 columns, space-separated, one node per line, without header (node_name float_value, e.g. "mynode_id 8.5"). There should be as many lines as nodes. At the moment, only one covariate is allowed. This will change in the next version.

Version

1.0.0

Requirements

• tensorflow
• keras
• ugtm
• networkx
• gensim
• numpy