SCNN
Note: A newer, cleaner implementation is now available at https://github.com/jcatw/dcnn.
An implementation of ~~search-~~diffusion-convolutional neural networks, a new model for graph-structured data.
Installation
Using pip:
pip install scnn
Usage
import numpy as np
from scnn import SCNN, data
from sklearn.metrics import f1_score
# Parse the cora dataset and return an adjacency matrix, a design matrix, and a 1-hot label matrix
A, X, Y = data.parse_cora()
# Construct array indices for the training, validation, and test sets
n_nodes = A.shape[0]
indices = np.arange(n_nodes)
train_indices = indices[:n_nodes // 3]
valid_indices = indices[n_nodes // 3:(2* n_nodes) // 3]
test_indices = indices[(2* n_nodes) // 3:]
# Instantiate an SCNN and fit it to cora
scnn = SCNN()
scnn.fit(A, X, Y, train_indices=train_indices, valid_indices=valid_indices)
# Predict labels for the test set
preds = scnn.predict(X, test_indices)
actuals = np.argmax(Y[test_indices,:], axis=1)
# Display performance
print 'F score: %.4f' % (f1_score(actuals, preds))
What's with the S?
Historical reasons - these were once called search-convolutional neural networks.