def fit_predict(self, features, adjacency_matrix):
adjacency_matrix = scipy.sparse.csr_matrix(adjacency_matrix)
nodes_number = adjacency_matrix.shape[0]
X = None
if self.X is None:
random_walker = gvnr.preprocessing.random_walker.RandomWalker(
adjacency_matrix,
walks_length=40,
walks_number=number_of_walks)
random_walks = random_walker.build_random_walks()
slider = gvnr.preprocessing.window_slider.WindowSlider(
random_walks,
nodes_number,
window_size=5,
window_factor="decreasing")
X = slider.build_cooccurrence_matrix()
else:
X = self.X
model = gvnr.models.gvnr.gvnr()
model.fit(X,
learn_rate=0.001,
embedding_size=80,
batch_size=batch_size,
n_epochs=2,
k_neg=num_neg,
x_min=x_min)
I = normalize(model.get_embeddings(embedding='I') +
model.get_embeddings(embedding='J'),
axis=0)
return I
def run(baselines, save=False):
logger.info("Running evaluations with parameters:")
for d in baselines:
logger.info(f"{d}: {baselines[d]}")
for dataset in baselines:
binary_vectors, tfidf_vectors, svd_vectors, adjacency_matrix, labels, gt_mask = gvnr.data.datasets.get_dataset(
dataset)
X_file = pkg_resources.resource_filename(
"gvnr", 'resources/{0}_X.npz'.format(dataset))
X = None
if os.path.isfile(X_file):
X = scipy.sparse.load_npz(X_file)
else:
random_walker = gvnr.preprocessing.random_walker.RandomWalker(
scipy.sparse.csr_matrix(adjacency_matrix),
walks_length=40,
walks_number=80)
random_walks = random_walker.build_random_walks()
slider = gvnr.preprocessing.window_slider.WindowSlider(
random_walks,
adjacency_matrix.shape[0],
window_size=5,
window_factor="decreasing")
X = slider.build_cooccurrence_matrix()
scipy.sparse.save_npz(X_file, X)
for baseline in baselines[dataset]:
model = None
features = None
if baseline == "binary":
model = gvnr.models.wrappers.binary_wrapper()
features = binary_vectors
if baseline == "tfidf":
model = gvnr.models.wrappers.tfidf_wrapper()
features = tfidf_vectors
if baseline == "svd":
model = gvnr.models.wrappers.svd_wrapper()
features = svd_vectors
if baseline == "tadw":
model = gvnr.models.wrappers.tadw_wrapper()
features = svd_vectors
if baseline == "netmf":
model = gvnr.models.wrappers.netmf_wrapper()
if baseline == "netmf_small":
model = gvnr.models.wrappers.netmf_small_wrapper()
if baseline == "deepwalk":
model = gvnr.models.wrappers.deepwalk_wrapper()
features = None
if baseline == "deepwalk_svd":
model = gvnr.models.wrappers.deepwalk_svd_wrapper()
features = svd_vectors
if baseline == "netmf_svd":
model = gvnr.models.wrappers.netmf_svd_wrapper()
features = svd_vectors
if baseline == "glove":
model = gvnr.models.wrappers.glove_wrapper(X=X)
features = binary_vectors
if baseline == "gvnr_no_filter":
model = gvnr.models.wrappers.gvnr_no_filter_wrapper(X=X)
features = binary_vectors
if baseline == "gvnr":
model = gvnr.models.wrappers.gvnr_wrapper(X=X)
features = binary_vectors
if baseline == "gvnrt":
model = gvnr.models.wrappers.gvnrt_wrapper(X=X)
features = binary_vectors
vectors = model.fit_predict(features, adjacency_matrix)
if save is True:
filename = os.path.abspath(
os.path.join(
os.path.dirname(__file__), os.path.pardir,
f"embeddings/{dataset}_{baseline}_vectors.csv"))
logger.info(f"Saving embeddings to {filename}...")
np.savetxt(filename, vectors)
vector = vectors[gt_mask]
scores = None
if dataset in ["cora", "aminer", "wiki", "citeseer"]:
scores = gvnr.evaluation.get_score(vectors, labels)
if dataset in ["flickr", "blogcatalog", "wikipedia", "ppi"]:
scores = gvnr.evaluation.get_score(vectors,
labels,
multilabels=True)
logger.info(
"DATASET: {0}, BASELINE: {1} => f1_micro {2} f1_macro {2}".
format(dataset, baseline, proportions))
logger.info(" ".join([
"&{0:.1f}".format(s * 100)
for s in list(scores["f1_micro"]) + list(scores["f1_macro"])
]))
labels: groundtruth labels for classification
gt_mask: mask to select nodes that are linked with a label (only useful with aminer)
"""
binary_vectors, tfidf_vectors, svd_vectors, adjacency_matrix, labels, gt_mask = gvnr.data.datasets.get_dataset(
"cora")
logger.info("2) PREPROCESSING THE DATA")
"""
We perform random walks on the network. N being the number of nodes, we get N*10 sequences of nodes of lengths 40.
These are stored in a numpy.ndarray
"""
random_walker = gvnr.preprocessing.random_walker.RandomWalker(
scipy.sparse.csr_matrix(adjacency_matrix),
walks_length=40,
walks_number=10)
random_walks = random_walker.build_random_walks()
"""
From these sequences of nodes, we slide a window to increment a matrix of counts of co-occurring nodes. We look 5
context nodes on the left and on the right of a target node. The window_factor is chosen such that co-occurrence counts
are decreasingly incremented given the distance to the target node e.g:
[1/5, 1/4, 1/3, 1/2, 1, target_node, 1, 1/2, 1/3, 1/4, 1/5]
"""
slider = gvnr.preprocessing.window_slider.WindowSlider(
random_walks,
adjacency_matrix.shape[0],
window_size=5,
window_factor="decreasing")
X = slider.build_cooccurrence_matrix()
logger.info("3) TRAINING THE MODEL")
"""