def initiEmbeddings():
if not os.path.exists(basepath + 'line_embeddings.emb'):
G = nx.read_gml(basepath + "network.gml", label=None)
model = LINE(G, embedding_size=128, order='second')
model.train(batch_size=1024, epochs=150, verbose=2)
embeddings = model.get_embeddings()
# print (embeddings)
t = []
for key in embeddings:
t.append(embeddings[key])
#print (embeddings[key])
np.savetxt(basepath + 'line_embeddings.emb', np.array(t))
# Emneddings = np.loadtxt(basepath + 'line_embeddings.emb')
Embeddings = np.loadtxt(basepath + 'line_embeddings.emb')
return Embeddings
def embedding_trainer(G,
embedder,
epochs=250,
seed=1234,
learning_rate=0.05,
embedding_dim=96,
batch_size=1024,
walk_length=30,
num_walks=200,
window=10,
p=1.0,
q=1.0,
workers=1,
temp_folder=None):
if embedder == 'node2vec':
if not os.path.isdir(temp_folder):
os.mkdir(temp_folder)
model = Node2Vec(G,
dimensions=embedding_dim,
walk_length=walk_length,
num_walks=num_walks,
p=p,
q=q,
weight_key='weight',
workers=5,
temp_folder=temp_folder)
model = model.fit(window=window,
min_count=1,
seed=seed,
alpha=learning_rate,
batch_words=4)
model.wv.save_word2vec_format('./temp_embeddings_file.emb')
embeddings = node2vec_embedder('temp_embeddings_file.emb')
os.remove('./temp_embeddings_file.emb')
elif embedder == 'line':
model = LINE(G, embedding_size=embedding_dim, order='second')
model.train(batch_size=batch_size, epochs=epochs, verbose=2)
embeddings = model.get_embeddings()
elif embedder == 'rolx':
# embeddings = np.load('./data/rolx_embeddings.npy')[()]
with open(args["directory+'/embeddings/rolx_embedding.json'"]) as fp:
embeddings = json.load(fp)
# embeddings = {str(k):v for k,v in embeddings.items()}
return embeddings
emb_list = []
for k in X:
emb_list.append(embeddings[k])
emb_list = np.array(emb_list)
model = TSNE(n_components=2)
node_pos = model.fit_transform(emb_list)
color_idx = {}
for i in range(len(X)):
color_idx.setdefault(Y[i][0], [])
color_idx[Y[i][0]].append(i)
for c, idx in color_idx.items():
plt.scatter(node_pos[idx, 0], node_pos[idx, 1], label=c)
plt.legend()
plt.show()
if __name__ == "__main__":
G = nx.read_edgelist('../data/wiki/Wiki_edgelist.txt',
create_using=nx.DiGraph(), nodetype=None, data=[('weight', int)])
model = LINE(G, embedding_size=128, order='second')
model.train(batch_size=1024, epochs=1, verbose=1)
embeddings = model.get_embeddings()
evaluate_embeddings(embeddings)
plot_embeddings(embeddings)
for k in X:
emb_list.append(embeddings[k])
emb_list = np.array(emb_list)
model = TSNE(n_components=2)
node_pos = model.fit_transform(emb_list)
color_idx = {}
for i in range(len(X)):
color_idx.setdefault(Y[i][0], [])
color_idx[Y[i][0]].append(i)
for c, idx in color_idx.items():
plt.scatter(node_pos[idx, 0], node_pos[idx, 1], label=c)
plt.legend()
plt.show()
if __name__ == "__main__":
G = nx.read_edgelist('../data/zachary/zachary_edgelist.txt',
create_using=nx.DiGraph(),
nodetype=None,
data=[('weight', int)])
model = LINE(G, embedding_size=128, order='second')
model.train(batch_size=1024, epochs=50, verbose=2)
embeddings = model.get_embeddings()
evaluate_embeddings(embeddings)
plot_embeddings(embeddings)
# model = LINE(G, embedding_size=config.embedding_size, order='second')
# elif config.model_name.lower() == 'struc2vec':
# model = Struc2Vec(G, 10, 80, workers=4, verbose=40, )
# elif config.model_name.lower() == 'sdne':
# model = SDNE(G, hidden_size=[256, 128], )
# elif config.model_name.lower() == 'node2vec':
# model = Node2Vec(G, walk_length=config.walk_length, num_walks=config.num_walks,p=0.25, q=4, workers=config.workers)
# else:
# model = DeepWalk(G, walk_length=config.walk_length, num_walks=config.num_walks, workers=config.workers)
# print('参数文件中模型名称错误，采用默认的deepwalk模型')
model = LINE(G,
embedding_size=config.embedding_size,
order=config.line_order,
negative_ratio=config.negative_ratio)
model.train(batch_size=config.line_batch_size, epochs=50, verbose=2)
# embeddings = model.get_embeddings()
# evaluate_embeddings(embeddings)
# plot_embeddings(embeddings)
# entity_dict = {}
# f = open('/Users/admin/Desktop/GraphEmbedding-deeplearning/data/XunYiWenYao/寻医问药category.txt','r',encoding='utf-8')
# for i in f.readlines():
# entity_dict[i.strip().split(' ')[0]] = i.strip().split(' ')[1]
#
# max_similarity = 0
# max_similarity_tuple = []
#
# for k1, v1 in entity_dict.items():
#
def embedding_feature(self, full_data, key='uid', target='task_id', embedding_size=16, epoch=10, window_size=5,
mode='LINE', suffix='cnt', order='second', graph='day'):
# 调用GraphEmbedding生成全局embedding
model_path = './yangzhe/model/n2v/{}_{}_{}_{}_{}_{}.pkl'.format(mode, suffix, key, target, graph,
embedding_size)
if not os.path.exists(model_path):
G = nx.read_edgelist('./yangzhe/feature/graph/{}_{}_{}_graph.csv'.format(target, suffix, graph),
create_using=nx.DiGraph(), nodetype=None, data=[('weight', int)])
tf.keras.backend.clear_session()
if mode == 'LINE':
model = LINE(graph=G, embedding_size=embedding_size, order=order)
model.train(batch_size=64, epochs=epoch, verbose=1)
else:
model = DeepWalk(G, walk_length=10, num_walks=80, workers=1)
model.train(embed_size=embedding_size, window_size=window_size, workers=5)
with open(model_path, 'wb') as f:
pickle.dump(model.get_embeddings(), f)
# LINE对应的一阶特征与二阶特征
if order == 'all':
embedding_size = embedding_size * 2
# 有些target的embedding是没有学习到的，这些不不存在于这个dict中，所以embedding中没有这些target所对应的行
with open(model_path, 'rb') as f:
embedding_dict = pickle.load(f)
embedding = pd.DataFrame()
embedding[target] = embedding_dict.keys()
embedding['embedding'] = [embedding_dict[i] for i in embedding[target].values]
embedding[target] = embedding[target].astype(int)
sentences = full_data[[key, target]].groupby([key])[target].agg(list)
# 这里是根据每个用户的历史曝光target进行均值来求用户的embedding，这些target应该在embedding[target]中
task_id_have_embedding = set(embedding[target])
lbl = LabelEncoder()
lbl.fit(embedding[target])
emb_matrix = np.array([i for i in embedding['embedding'].values])
emb_mean = []
for idx_list in sentences.values.tolist():
need_key = [x for x in idx_list if x in task_id_have_embedding]
if len(need_key) == 0:
mean = np.zeros((embedding_size,))
else:
index = lbl.transform(need_key)
mean = np.mean(emb_matrix[index], axis=0)
emb_mean.append(mean)
emb_feature = np.asarray(emb_mean)
mean_col = ['{}_{}(MainKEY)_{}_MEAN_Window{}_{}'.format(mode, key, target, window_size, i) for i in
range(embedding_size)]
emb_feature = pd.DataFrame(emb_feature, columns=mean_col)
emb_feature[key] = sentences.index
# target对应的embedding矩阵也存起来
embeddings = np.concatenate(embedding['embedding'].values).reshape(-1, embedding_size)
embeddings = pd.DataFrame(embeddings,
columns=["{}_{}_{}(MainKEY)_Window{}_{}".format(mode, key, target, window_size, i)
for i in range(embedding_size)])
embedding[embeddings.columns] = embeddings
del embedding['embedding']
return emb_feature.reset_index(drop=True), embedding.reset_index(drop=True)