def _add_ndata(self):
vectorizer = CountVectorizer(min_df=5)
features = vectorizer.fit_transform(
self.data['plot_keywords'].fillna('').values)
self.g.nodes['movie'].data['feat'] = torch.from_numpy(
features.toarray()).float()
self.g.nodes['movie'].data['label'] = torch.from_numpy(
self.labels).long()
# actor和director顶点的特征为其关联的movie顶点特征的平均
self.g.multi_update_all(
{
'ma': (fn.copy_u('feat', 'm'), fn.mean('m', 'feat')),
'md': (fn.copy_u('feat', 'm'), fn.mean('m', 'feat'))
}, 'sum')
n_movies = len(self.movies)
train_idx, val_idx, test_idx = split_idx(np.arange(n_movies), 400, 400,
self._seed)
self.g.nodes['movie'].data['train_mask'] = generate_mask_tensor(
idx2mask(train_idx, n_movies))
self.g.nodes['movie'].data['val_mask'] = generate_mask_tensor(
idx2mask(val_idx, n_movies))
self.g.nodes['movie'].data['test_mask'] = generate_mask_tensor(
idx2mask(test_idx, n_movies))
def process(self):
with open(os.path.join(self.raw_dir, 'ACM3025.pkl'), 'rb') as f:
data = pickle.load(f)
features = torch.from_numpy(
data['feature'].todense()).float() # (3025, 1870)
labels = torch.from_numpy(
data['label'].todense()).long().nonzero(as_tuple=True)[1] # (3025)
# Adjacency matrices for meta-path based neighbors
# (Mufei): I verified both of them are binary adjacency matrices with self loops
author_g = dgl.from_scipy(data['PAP'])
subject_g = dgl.from_scipy(data['PLP'])
self.gs = [author_g, subject_g]
num_nodes = data['label'].shape[0]
train_mask = generate_mask_tensor(
idx2mask(data['train_idx'][0], num_nodes))
val_mask = generate_mask_tensor(idx2mask(data['val_idx'][0],
num_nodes))
test_mask = generate_mask_tensor(
idx2mask(data['test_idx'][0], num_nodes))
for g in self.gs:
g.ndata['feat'] = features
g.ndata['label'] = labels
g.ndata['train_mask'] = train_mask
g.ndata['val_mask'] = val_mask
g.ndata['test_mask'] = test_mask
def _add_ndata(self):
_raw_file2 = os.path.join(self.raw_dir, 'DBLP4057_GAT_with_idx.mat')
if not os.path.exists(_raw_file2):
raise FileNotFoundError('请手动下载文件 {} 提取码：6b3h 并保存到 {}'.format(
self._url2, _raw_file2
))
mat = sio.loadmat(_raw_file2)
self.g.nodes['author'].data['feat'] = torch.from_numpy(mat['features']).float()
self.g.nodes['author'].data['label'] = torch.tensor(self.authors['label'].to_list())
n_authors = len(self.authors)
train_idx, val_idx, test_idx = split_idx(np.arange(n_authors), 800, 400, self._seed)
self.g.nodes['author'].data['train_mask'] = generate_mask_tensor(idx2mask(train_idx, n_authors))
self.g.nodes['author'].data['val_mask'] = generate_mask_tensor(idx2mask(val_idx, n_authors))
self.g.nodes['author'].data['test_mask'] = generate_mask_tensor(idx2mask(test_idx, n_authors))
self.g.nodes['conf'].data['label'] = torch.tensor(self.confs['label'].to_list())
def process(self):
data = sio.loadmat(os.path.join(self.raw_dir, 'DBLP4057_GAT_with_idx.mat'))
apa_g = dgl.graph(data['net_APA'].nonzero())
apcpa_g = dgl.graph(data['net_APCPA'].nonzero())
aptpa_g = dgl.graph(data['net_APTPA'].nonzero())
self.gs = [apa_g, apcpa_g, aptpa_g]
features = torch.from_numpy(data['features']).float()
labels = torch.from_numpy(data['label'].nonzero()[1])
num_nodes = data['label'].shape[0]
train_mask = generate_mask_tensor(idx2mask(data['train_idx'][0], num_nodes))
val_mask = generate_mask_tensor(idx2mask(data['val_idx'][0], num_nodes))
test_mask = generate_mask_tensor(idx2mask(data['test_idx'][0], num_nodes))
for g in self.gs:
g.ndata['feat'] = features
g.ndata['label'] = labels
g.ndata['train_mask'] = train_mask
g.ndata['val_mask'] = val_mask
g.ndata['test_mask'] = test_mask
def process(self):
data = sio.loadmat(os.path.join(self.raw_dir, 'imdb5k.mat'))
mam_g = dgl.graph(data['MAM'].nonzero())
mdm_g = dgl.graph(data['MDM'].nonzero())
# mym_g = dgl.graph(data['MYM'].nonzero())
self.gs = [mam_g, mdm_g]
features = torch.from_numpy(data['feature']).float()
num_nodes = features.shape[0]
labels = torch.full((num_nodes, ), -1, dtype=torch.long)
idx, label = data['label'].nonzero()
labels[idx] = torch.from_numpy(label)
train_mask = generate_mask_tensor(
idx2mask(data['train_idx'][0], num_nodes))
val_mask = generate_mask_tensor(idx2mask(data['val_idx'][0],
num_nodes))
test_mask = generate_mask_tensor(
idx2mask(data['test_idx'][0], num_nodes))
for g in self.gs:
g.ndata['feat'] = features
g.ndata['label'] = labels
g.ndata['train_mask'] = train_mask
g.ndata['val_mask'] = val_mask
g.ndata['test_mask'] = test_mask
def process(self):
self.g = dgl.heterograph(self._read_edges())
feats = self._read_feats()
for ntype, feat in feats.items():
self.g.nodes[ntype].data['feat'] = feat
labels = torch.from_numpy(
np.load(os.path.join(self.raw_path, 'labels.npy'))).long()
self._num_classes = labels.max().item() + 1
self.g.nodes[self.predict_ntype].data['label'] = labels
n = self.g.num_nodes(self.predict_ntype)
for split in ('train', 'val', 'test'):
idx = np.load(os.path.join(self.raw_path, f'{split}_60.npy'))
mask = generate_mask_tensor(idx2mask(idx, n))
self.g.nodes[self.predict_ntype].data[f'{split}_mask'] = mask
pos_i, pos_j = sp.load_npz(os.path.join(self.raw_path,
'pos.npz')).nonzero()
self.pos_i, self.pos_j = torch.from_numpy(
pos_i).long(), torch.from_numpy(pos_j).long()
def process(self):
data = sio.loadmat(os.path.join(self.raw_dir, 'ACM.mat'))
p_vs_l = data['PvsL'] # paper-field?
p_vs_a = data['PvsA'] # paper-author
p_vs_t = data['PvsT'] # paper-term, bag of words
p_vs_c = data['PvsC'] # paper-conference, labels come from that
# We assign
# (1) KDD papers as class 0 (data mining),
# (2) SIGMOD and VLDB papers as class 1 (database),
# (3) SIGCOMM and MobiCOMM papers as class 2 (communication)
conf_ids = [0, 1, 9, 10, 13]
label_ids = [0, 1, 2, 2, 1]
p_vs_c_filter = p_vs_c[:, conf_ids]
p_selected = (p_vs_c_filter.sum(1) != 0).A1.nonzero()[0]
p_vs_l = p_vs_l[p_selected]
p_vs_a = p_vs_a[p_selected]
p_vs_t = p_vs_t[p_selected]
p_vs_c = p_vs_c[p_selected]
self.g = dgl.heterograph({
('paper', 'pa', 'author'):
p_vs_a.nonzero(),
('author', 'ap', 'paper'):
p_vs_a.transpose().nonzero(),
('paper', 'pf', 'field'):
p_vs_l.nonzero(),
('field', 'fp', 'paper'):
p_vs_l.transpose().nonzero()
})
paper_features = torch.FloatTensor(p_vs_t.toarray()) # (4025, 1903)
pc_p, pc_c = p_vs_c.nonzero()
paper_labels = np.zeros(len(p_selected), dtype=np.int64)
for conf_id, label_id in zip(conf_ids, label_ids):
paper_labels[pc_p[pc_c == conf_id]] = label_id
paper_labels = torch.from_numpy(paper_labels)
float_mask = np.zeros(len(pc_p))
for conf_id in conf_ids:
pc_c_mask = (pc_c == conf_id)
float_mask[pc_c_mask] = np.random.permutation(
np.linspace(0, 1, pc_c_mask.sum()))
train_idx = np.where(float_mask <= 0.2)[0]
val_idx = np.where((float_mask > 0.2) & (float_mask <= 0.3))[0]
test_idx = np.where(float_mask > 0.3)[0]
num_paper_nodes = self.g.num_nodes('paper')
train_mask = generate_mask_tensor(idx2mask(train_idx, num_paper_nodes))
val_mask = generate_mask_tensor(idx2mask(val_idx, num_paper_nodes))
test_mask = generate_mask_tensor(idx2mask(test_idx, num_paper_nodes))
self.g.nodes['paper'].data['feat'] = paper_features
self.g.nodes['paper'].data['label'] = paper_labels
self.g.nodes['paper'].data['train_mask'] = train_mask
self.g.nodes['paper'].data['val_mask'] = val_mask
self.g.nodes['paper'].data['test_mask'] = test_mask
# author顶点的特征为其关联的paper顶点特征的平均
self.g.multi_update_all(
{'pa': (fn.copy_u('feat', 'm'), fn.mean('m', 'feat'))}, 'sum')
self.g.nodes['field'].data['feat'] = torch.eye(
self.g.num_nodes('field'))