import numpy as np
import pandas as pd
import spacy
from datetime import datetime
import torch
from typing import Any, Dict, List
import math
import os
import dgl
import dgl.function as fn
import gc
from itertools import combinations
from scipy import sparse
from scipy.sparse import coo_matrix
from scipy.sparse import csr_matrix
from sklearn import metrics
from sklearn.cluster import KMeans, DBSCAN
from torch.utils.data import Dataset
from torch.functional import Tensor
from torch import nn
import torch.nn.functional as F
from torch_geometric.nn import GATConv
from torch.nn import Linear, BatchNorm1d, Sequential, ModuleList, ReLU, Dropout
from torch_geometric.data import Data
from torch_geometric.loader import NeighborSampler
import random
import argparse
from time import localtime, strftime, time
import networkx as nx
import json
import torch.optim as optim
import sys
sys.path.append(os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
[docs]class FinEvent:
r"""The FinEvent model for social event detection that uses graph neural networks
and reinforcement learning for adaptive event detection.
.. note::
This detector uses graph neural networks and reinforcement learning to identify events in social media data.
The model requires a dataset object with a load_data() method.
Parameters
----------
dataset : object
The dataset object containing social media data.
Must provide load_data() method that returns the raw data.
n_epochs : int, optional
Number of training epochs. Default: ``1``.
window_size : int, optional
Size of sliding window for incremental learning. Default: ``3``.
patience : int, optional
Number of epochs to wait before early stopping. Default: ``5``.
margin : float, optional
Margin for triplet loss. Default: ``3.0``.
lr : float, optional
Learning rate. Default: ``1e-3``.
batch_size : int, optional
Mini-batch size. Default: ``50``.
hidden_dim : int, optional
Hidden layer dimension. Default: ``128``.
out_dim : int, optional
Output dimension. Default: ``64``.
heads : int, optional
Number of attention heads. Default: ``4``.
validation_percent : float, optional
Percentage of data for validation. Default: ``0.2``.
use_hardest_neg : bool, optional
Whether to use hardest negative mining. Default: ``False``.
is_shared : bool, optional
Whether to use shared parameters. Default: ``False``.
inter_opt : str, optional
Integration option for multi-view features. Default: ``'cat_w_avg'``.
is_initial : bool, optional
Whether to initialize model. Default: ``True``.
sampler : str, optional
Type of sampler to use. Default: ``'RL_sampler'``.
cluster_type : str, optional
Clustering algorithm to use. Default: ``'kmeans'``.
threshold_start0 : list, optional
Initial thresholds for RL-0. Default: ``[[0.2], [0.2], [0.2]]``.
RL_step0 : float, optional
Step size for RL-0. Default: ``0.02``.
RL_start0 : int, optional
Starting point for RL-0. Default: ``0``.
eps_start : float, optional
Initial epsilon for RL-1. Default: ``0.001``.
eps_step : float, optional
Step size for epsilon in RL-1. Default: ``0.02``.
min_Pts_start : int, optional
Initial minimum points for RL-1. Default: ``2``.
min_Pts_step : int, optional
Step size for minimum points in RL-1. Default: ``1``.
use_cuda : bool, optional
Whether to use GPU acceleration. Default: ``True``.
data_path : str, optional
Path to data directory. Default: ``'../model/model_saved/finevent/incremental_test/'``.
file_path : str, optional
Path to save model files. Default: ``'../model/model_saved/finevent/'``.
mask_path : str, optional
Path to attention mask file. Default: ``None``.
log_interval : int, optional
Number of steps between logging. Default: ``10``.
"""
def __init__(self,
# Hyper-parameters
dataset,
n_epochs=1,
window_size=3,
patience=5,
margin=3.0,
lr=1e-3,
batch_size=50,
hidden_dim=128,
out_dim=64,
heads=4,
validation_percent=0.2,
use_hardest_neg=False,
is_shared=False,
inter_opt='cat_w_avg',
is_initial=True,
sampler='RL_sampler',
cluster_type='kmeans',
# RL-0
threshold_start0=[[0.2], [0.2], [0.2]],
RL_step0=0.02,
RL_start0=0,
# RL-1
eps_start=0.001,
eps_step=0.02,
min_Pts_start=2,
min_Pts_step=1,
# Other arguments
use_cuda=True,
data_path='../model/model_saved/finevent/incremental_test/',
file_path='../model/model_saved/finevent/',
mask_path=None,
log_interval=10):
self.dataset = dataset
# Hyper-parameters
self.n_epochs = n_epochs
self.window_size = window_size
self.patience = patience
self.margin = margin
self.lr = lr
self.batch_size = batch_size
self.hidden_dim = hidden_dim
self.out_dim = out_dim
self.heads = heads
self.validation_percent = validation_percent
self.use_hardest_neg = use_hardest_neg
self.is_shared = is_shared
self.inter_opt = inter_opt
self.is_initial = is_initial
self.sampler = sampler
self.cluster_type = cluster_type
# RL-0
self.threshold_start0 = threshold_start0
self.RL_step0 = RL_step0
self.RL_start0 = RL_start0
# RL-1
self.eps_start = eps_start
self.eps_step = eps_step
self.min_Pts_start = min_Pts_start
self.min_Pts_step = min_Pts_step
# Other arguments
self.use_cuda = use_cuda
self.data_path = data_path
self.file_path = file_path
self.mask_path = mask_path
self.log_interval = log_interval
[docs] def preprocess(self):
preprocessor = Preprocessor()
preprocessor.generate_initial_features(self.dataset)
preprocessor.construct_graph(self.dataset)
preprocessor.save_edge_index()
[docs] def fit(self):
args=self
embedding_save_path = args.data_path + '/embeddings'
os.makedirs(embedding_save_path, exist_ok=True)
print('embedding save path: ', embedding_save_path)
# record hyper-parameters
# with open(embedding_save_path + '/args.txt', 'w') as f:
# json.dump(args.__dict__, f, indent=2)
print('Batch Size:', args.batch_size)
print('Intra Agg Mode:', args.is_shared)
print('Inter Agg Mode:', args.inter_opt)
print('Reserve node config?', args.is_initial)
data_split = np.load(args.data_path + '/data_split.npy')
if args.use_hardest_neg:
loss_fn = OnlineTripletLoss(args.margin, HardestNegativeTripletSelector(args.margin))
else:
loss_fn = OnlineTripletLoss(args.margin, RandomNegativeTripletSelector(args.margin))
# define metrics
BCL_metrics = [AverageNonzeroTripletsMetric()]
# define detection stage
Streaming = FinEvent_model(args)
# pre-train stage: train on initial graph
train_i = 0
self.model, self.RL_thresholds = Streaming.initial_maintain(train_i=train_i,
i=0,
metrics=BCL_metrics,
embedding_save_path=embedding_save_path,
loss_fn=loss_fn,
model=None)
# detection-maintenance stage: incremental training and detection
for i in range(1, data_split.shape[0]):
# infer every block
self.model = Streaming.inference(train_i=train_i,
i=i,
metrics=BCL_metrics,
embedding_save_path=embedding_save_path,
loss_fn=loss_fn,
model=self.model,
RL_thresholds=self.RL_thresholds)
# maintenance in window size and desert the last block
if i % args.window_size == 0 and i != data_split.shape[0] - 1:
train_i = i
self.model, self.RL_thresholds = Streaming.initial_maintain(train_i=train_i,
i=i,
metrics=BCL_metrics,
embedding_save_path=embedding_save_path,
loss_fn=loss_fn,
model=None)
[docs] def detection(self):
args=self
"""
:param eval_data_path: Path to the detection data
:param eval_metrics: List of detection metrics
:param embedding_save_path: Path to save embeddings if needed
:param best_model_path: Path to the best trained model
:param loss_fn: Loss function used during detection
:return: None
"""
start_time = time()
# Load detection data
print("Loading detection data...")
relation_ids = ['entity', 'userid', 'word']
homo_data = create_offline_homodataset(args.data_path, [0, 0])
multi_r_data = create_multi_relational_graph(args.data_path, relation_ids, [0, 0])
print("detection data loaded. Time elapsed: {:.2f} seconds".format(time() - start_time))
device = torch.device('cuda' if torch.cuda.is_available() and args.use_cuda else 'cpu')
# Load the best trained model
print("Loading the best trained model...")
best_model_path = args.data_path + 'embeddings/block_0/models/best.pt'
feat_dim = homo_data.x.size(1)
num_relations = len(multi_r_data)
self.model = MarGNN((feat_dim, args.hidden_dim, args.out_dim, args.heads),
num_relations=num_relations, inter_opt=args.inter_opt, is_shared=args.is_shared)
state_dict = torch.load(best_model_path)
self.model.load_state_dict(state_dict)
self.model.to(device) # 将模型移动到指定设备（如果使用GPU）
# 设置模型为评估模式
self.model.eval()
print("Best model loaded and set to eval mode. Time elapsed: {:.2f} seconds".format(time() - start_time))
RL_thresholds = torch.FloatTensor(args.threshold_start0)
filtered_multi_r_data = torch.load('../model/model_saved/finevent/multi_remain_data.pt')
# Sampling nodes
print("Sampling nodes...")
sampler = MySampler(args.sampler)
test_num_samples = homo_data.test_mask.size(0)
num_batches = int(test_num_samples / args.batch_size) + 1
extract_features = []
for batch in range(num_batches):
print(f"Processing batch {batch + 1}/{num_batches}...")
i_start = args.batch_size * batch
i_end = min((batch + 1) * args.batch_size, test_num_samples)
batch_nodes = homo_data.test_mask[i_start:i_end]
batch_labels = homo_data.y[batch_nodes]
adjs, n_ids = sampler.sample(filtered_multi_r_data, node_idx=batch_nodes, sizes=[-1, -1],
batch_size=args.batch_size)
# Perform prediction
with torch.no_grad():
pred = self.model(homo_data.x, adjs, n_ids, device, RL_thresholds)
extract_features.append(pred.cpu().detach())
print(f"Batch {batch + 1} processed.")
extract_features = torch.cat(extract_features, dim=0)
all_nodes = homo_data.test_mask
ground_truths = homo_data.y[all_nodes]
X = extract_features.cpu().detach().numpy()
assert ground_truths.shape[0] == X.shape[0]
# Get the total number of classes
n_classes = len(set(ground_truths.tolist()))
# k-means clustering
print("Performing k-means clustering...")
kmeans = KMeans(n_clusters=n_classes, random_state=0).fit(X)
predictions = kmeans.labels_
print("k-means clustering done. Time elapsed: {:.2f} seconds".format(time() - start_time))
print("Detection complete. Total time elapsed: {:.2f} seconds".format(time() - start_time))
return ground_truths, predictions
[docs] def evaluate(self, predictions, ground_truths):
ars = metrics.adjusted_rand_score(ground_truths, predictions)
# Calculate Adjusted Mutual Information (AMI)
ami = metrics.adjusted_mutual_info_score(ground_truths, predictions)
# Calculate Normalized Mutual Information (NMI)
nmi = metrics.normalized_mutual_info_score(ground_truths, predictions)
print(f"Model Adjusted Rand Index (ARI): {ars}")
print(f"Model Adjusted Mutual Information (AMI): {ami}")
print(f"Model Normalized Mutual Information (NMI): {nmi}")
return ars, ami, nmi
[docs]class FinEvent_model(FinEvent):
def __init__(self, args) -> None:
# register args
super().__init__(dataset="")
self.args = args
[docs] def inference(self,
train_i, i,
metrics,
embedding_save_path,
loss_fn,
model,
RL_thresholds=None,
loss_fn_dgi=None):
model = MarGNN()
# make dir for graph i
# ./incremental_0808//embeddings_0403005348/block_xxx
save_path_i = embedding_save_path + '/block_' + str(i)
if not os.path.isdir(save_path_i):
os.mkdir(save_path_i)
# load data
relation_ids: List[str] = ['entity', 'userid', 'word']
homo_data = create_homodataset(self.args.data_path, [train_i, i], self.args.validation_percent)
multi_r_data = create_multi_relational_graph(self.args.data_path, relation_ids, [train_i, i])
print('embedding save path: ', embedding_save_path)
num_relations = len(multi_r_data)
device = torch.device('cuda:0' if torch.cuda.is_available() and self.args.use_cuda else 'cpu')
# input dimension (300 in our paper)
features = homo_data.x
feat_dim = features.size(1)
# prepare graph configs for node filtering
if self.args.is_initial:
print('prepare node configures...')
pre_node_dist(multi_r_data, homo_data.x, save_path_i)
filter_path = save_path_i
else:
filter_path = save_path_i
if model is None:
assert 'Cannot find pre-trained model'
# directly predict
message = "\n------------ Directly predict on block " + str(i) + " ------------\n"
print(message)
print('RL Threshold using in this block:', RL_thresholds)
model.eval()
test_indices, labels = homo_data.test_mask, homo_data.y
test_num_samples = test_indices.size(0)
sampler = MySampler(self.args.sampler)
# filter neighbor in advance to fit with neighbor sampling
filtered_multi_r_data = RL_neighbor_filter(self,multi_r_data, RL_thresholds,
filter_path) if RL_thresholds is not None and self.args.sampler == 'RL_sampler' else multi_r_data
# batch testing
extract_features = torch.FloatTensor([])
num_batches = int(test_num_samples / self.args.batch_size) + 1
with torch.no_grad():
for batch in range(num_batches):
start_batch = time()
# split batch
i_start = self.args.batch_size * batch
i_end = min((batch + 1) * self.args.batch_size, test_num_samples)
batch_nodes = test_indices[i_start:i_end]
if not len(batch_nodes):
continue
# sampling neighbors of batch nodes
adjs, n_ids = sampler.sample(filtered_multi_r_data, node_idx=batch_nodes, sizes=[-1, -1],
batch_size=self.args.batch_size)
pred = model(homo_data.x, adjs, n_ids, device, RL_thresholds)
batch_seconds_spent = time() - start_batch
# for we haven't shuffle the test indices(see utils.py),
# the output embeddings can be simply stacked together
extract_features = torch.cat((extract_features, pred.cpu().detach()), dim=0)
del pred
gc.collect()
nmi, ami, ari, = evaluate_model(extract_features,
labels,
indices=test_indices,
epoch=-1, # just for test
num_isolated_nodes=0,
save_path=save_path_i,
is_validation=False,
cluster_type=self.args.cluster_type,
)
k_score = {"NMI": nmi, "AMI": ami, "ARI": ari}
del homo_data, multi_r_data, features, filtered_multi_r_data
torch.cuda.empty_cache()
return model, k_score
# train on initial/maintenance graphs, t == 0 or t % window_size == 0 in this paper
[docs] def initial_maintain(self,
train_i, i,
metrics,
embedding_save_path,
loss_fn,
model=None,
loss_fn_dgi=None):
"""
:param i:
:param data_split:
:param metrics:
:param embedding_save_path:
:param loss_fn:
:param model:
:param loss_fn_dgi:
:return:
"""
# make dir for graph i
# ./incremental_0808//embeddings_0403005348/block_xxx
save_path_i = embedding_save_path + '/block_' + str(i)
if not os.path.isdir(save_path_i):
os.mkdir(save_path_i)
# load data
relation_ids: List[str] = ['entity', 'userid', 'word']
homo_data = create_homodataset(self.args.data_path, [train_i, i], self.args.validation_percent)
multi_r_data = create_multi_relational_graph(self.args.data_path, relation_ids, [train_i, i])
num_relations = len(multi_r_data)
device = torch.device('cuda' if torch.cuda.is_available() and self.args.use_cuda else 'cpu')
# input dimension (300 in our paper)
num_dim = homo_data.x.size(0)
feat_dim = homo_data.x.size(1)
# prepare graph configs for node filtering
if self.args.is_initial:
print('prepare node configures...')
pre_node_dist(multi_r_data, homo_data.x, save_path_i)
filter_path = save_path_i
else:
filter_path = self.args.data_path + str(i)
if model is None: # pre-training stage in our paper
# print('Pre-Train Stage...')
model = MarGNN((feat_dim, self.args.hidden_dim, self.args.out_dim, self.args.heads),
num_relations=num_relations, inter_opt=self.args.inter_opt, is_shared=self.args.is_shared)
# define sampler
sampler = MySampler(self.args.sampler)
# load model to device
model.to(device)
# initialize RL thresholds
# RL_threshold: [[.5], [.5], [.5]]
RL_thresholds = torch.FloatTensor(self.args.threshold_start0)
# define optimizer
optimizer = optim.Adam(model.parameters(), lr=self.args.lr, weight_decay=1e-4)
# record training log
message = "\n------------ Start initial training / maintaining using block " + str(i) + " ------------\n"
print(message)
with open(save_path_i + '/log.txt', 'a') as f:
f.write(message)
# record the highest validation nmi ever got for early stopping
best_vali_nmi = 1e-9
best_epoch = 0
wait = 0
# record validation nmi of all epochs before early stop
all_vali_nmi = []
# record the time spent in seconds on each batch of all training/maintaining epochs
seconds_train_batches = []
# record the time spent in mins on each epoch
mins_train_epochs = []
# step13: start training
for epoch in range(self.args.n_epochs):
start_epoch = time()
losses = []
total_loss = 0.0
for metric in metrics:
metric.reset()
# Multi-Agent
# filter neighbor in advance to fit with neighbor sampling
filtered_multi_r_data = RL_neighbor_filter(self ,multi_r_data, RL_thresholds, filter_path) if epoch >= self.args.RL_start0 and self.args.sampler == 'RL_sampler' else multi_r_data
#filtered_multi_r_data = torch.load(self.args.file_path + 'multi_remain_data.pt')
print(f"Epoch {epoch + 1}/{self.args.n_epochs} - Starting training...")
model.train()
train_num_samples, valid_num_samples = homo_data.train_mask.size(0), homo_data.val_mask.size(0)
#train_num_samples, valid_num_samples = homo_data.train_mask.size(0) // 100, homo_data.val_mask.size(0)
all_num_samples = train_num_samples + valid_num_samples
# batch training
num_batches = int(train_num_samples / self.args.batch_size) + 1
for batch in range(num_batches):
start_batch = time()
# split batch
i_start = self.args.batch_size * batch
i_end = min((batch + 1) * self.args.batch_size, train_num_samples)
batch_nodes = homo_data.train_mask[i_start:i_end]
batch_labels = homo_data.y[batch_nodes]
print(
f"Epoch {epoch + 1}/{self.args.n_epochs} - Batch {batch + 1}/{num_batches}: Processing nodes {i_start} to {i_end}...")
# sampling neighbors of batch nodes
adjs, n_ids = sampler.sample(filtered_multi_r_data, node_idx=batch_nodes, sizes=[-1, -1],
batch_size=self.args.batch_size)
optimizer.zero_grad()
pred = model(homo_data.x, adjs, n_ids, device, RL_thresholds)
loss_outputs = loss_fn(pred, batch_labels)
loss = loss_outputs[0] if type(loss_outputs) in (tuple, list) else loss_outputs
losses.append(loss.item())
total_loss += loss.item()
for metric in metrics:
metric(pred, batch_labels, loss_outputs)
if batch % self.args.log_interval == 0:
message = 'Train: [{}/{} ({:.0f}%)]\tLoss: {:.6f}'.format(batch * self.args.batch_size,
train_num_samples, 100. * batch / ((
train_num_samples // self.args.batch_size) + 1),
np.mean(losses))
for metric in metrics:
message += '\t{}: {:.4f}'.format(metric.name(), metric.value())
print(message) # 输出到控制台
with open(save_path_i + '/log.txt', 'a') as f:
f.write(message)
losses = []
del pred, loss_outputs
gc.collect()
print(
f"Epoch {epoch + 1}/{self.args.n_epochs} - Batch {batch + 1}/{num_batches}: Performing backward pass...")
loss.backward()
optimizer.step()
batch_seconds_spent = time() - start_batch
seconds_train_batches.append(batch_seconds_spent)
del loss
gc.collect()
# step14: print loss
total_loss /= (batch + 1)
message = 'Epoch: {}/{}. Average loss: {:.4f}'.format(epoch + 1, self.args.n_epochs, total_loss)
for metric in metrics:
message += '\t{}: {:.4f}'.format(metric.name(), metric.value())
mins_spent = (time() - start_epoch) / 60
message += '\nThis epoch took {:.2f} mins'.format(mins_spent)
message += '\n'
print(message)
with open(save_path_i + '/log.txt', 'a') as f:
f.write(message)
mins_train_epochs.append(mins_spent)
# validation
# infer the representations of all tweets
model.eval()
# we recommand to forward all nodes and select the validation indices instead
extract_features = torch.FloatTensor([])
num_batches = int(all_num_samples / self.args.batch_size) + 1
# all mask are then splited into mini-batch in order
all_mask = torch.arange(0, num_dim, dtype=torch.long)
for batch in range(num_batches):
start_batch = time()
# split batch
i_start = self.args.batch_size * batch
i_end = min((batch + 1) * self.args.batch_size, all_num_samples)
batch_nodes = all_mask[i_start:i_end]
batch_labels = homo_data.y[batch_nodes]
# sampling neighbors of batch nodes
adjs, n_ids = sampler.sample(filtered_multi_r_data, node_idx=batch_nodes, sizes=[-1, -1],
batch_size=self.args.batch_size)
pred = model(homo_data.x, adjs, n_ids, device, RL_thresholds)
extract_features = torch.cat((extract_features, pred.cpu().detach()), dim=0)
del pred
gc.collect()
# save_embeddings(extract_features, save_path_i)
epoch = epoch + 1
# evaluate the model: conduct kMeans clustering on the validation and report NMI
validation_nmi = evaluate_model(extract_features[homo_data.val_mask],
homo_data.y,
indices=homo_data.val_mask,
epoch=epoch,
num_isolated_nodes=0,
save_path=save_path_i,
is_validation=True,
cluster_type=self.args.cluster_type)
all_vali_nmi.append(validation_nmi)
# step16: early stop
if validation_nmi > best_vali_nmi:
best_vali_nmi = validation_nmi
best_epoch = epoch
wait = 0
# save model
model_path = save_path_i + '/models'
if (epoch == 1) and (not os.path.isdir(model_path)):
os.mkdir(model_path)
p = model_path + '/best.pt'
torch.save(model.state_dict(), p)
print('Best model saved after epoch ', str(epoch))
else:
wait += 1
if wait >= self.args.patience:
print('Saved all_mins_spent')
print('Early stopping at epoch ', str(epoch))
print('Best model was at epoch ', str(best_epoch))
break
# end one epoch
# save all validation nmi
np.save(save_path_i + '/all_vali_nmi.npy', np.asarray(all_vali_nmi))
# save time spent on epochs
np.save(save_path_i + '/mins_train_epochs.npy', np.asarray(mins_train_epochs))
print('Saved mins_train_epochs.')
# save time spent on batches
np.save(save_path_i + '/seconds_train_batches.npy', np.asarray(seconds_train_batches))
print('Saved seconds_train_batches.')
# load the best model of the current block
best_model_path = save_path_i + '/models/best.pt'
model.load_state_dict(torch.load(best_model_path))
print("Best model loaded.")
del homo_data, multi_r_data
torch.cuda.empty_cache()
return model, RL_thresholds
[docs]class Preprocessor:
def __init__(self):
pass
[docs] def documents_to_features(self, df):
self.nlp = spacy.load("en_core_web_lg")
features = df.filtered_words.apply(lambda x: self.nlp(' '.join(x)).vector).values
return np.stack(features, axis=0)
[docs] def df_to_t_features(self, df):
t_features = np.asarray([self.extract_time_feature(t_str) for t_str in df['created_at']])
return t_features
[docs] def generate_initial_features(self, df, save_path='../model/model_saved/finevent/'):
os.makedirs(save_path, exist_ok=True)
print(df.shape)
print(df.head(10))
d_features = self.documents_to_features(df)
print("Document features generated.")
t_features = self.df_to_t_features(df)
print("Time features generated.")
combined_features = np.concatenate((d_features, t_features), axis=1)
print("Concatenated document features and time features.")
np.save(save_path + 'features_69612_0709_spacy_lg_zero_multiclasses_filtered.npy', combined_features)
print("Initial features saved.")
combined_features = np.load(save_path + 'features_69612_0709_spacy_lg_zero_multiclasses_filtered.npy')
print("Initial features loaded.")
print(combined_features.shape)
[docs] def construct_graph_from_df(self, df, G=None):
if G is None:
G = nx.Graph()
for _, row in df.iterrows():
tid = 't_' + str(row['tweet_id'])
G.add_node(tid)
G.nodes[tid]['tweet_id'] = True # right-hand side value is irrelevant for the lookup
user_ids = row['user_mentions']
user_ids.append(row['user_id'])
user_ids = ['u_' + str(each) for each in user_ids]
G.add_nodes_from(user_ids)
for each in user_ids:
G.nodes[each]['user_id'] = True
entities = row['entities']
G.add_nodes_from(entities)
for each in entities:
G.nodes[each]['entity'] = True
words = row['filtered_words']
words = ['w_' + each for each in words]
G.add_nodes_from(words)
for each in words:
G.nodes[each]['word'] = True
edges = []
edges += [(tid, each) for each in user_ids]
edges += [(tid, each) for each in entities]
edges += [(tid, each) for each in words]
G.add_edges_from(edges)
return G
[docs] def construct_incremental_dataset(self, df, save_path, features, test=True):
test_ini_size = 500
test_incr_size = 100
data_split = []
all_graph_mins = []
message = ""
distinct_dates = df.date.unique()
print("Number of distinct dates: ", len(distinct_dates))
print()
message += "Number of distinct dates: "
message += str(len(distinct_dates))
message += "\n"
print("Start constructing initial graph ...")
message += "\nStart constructing initial graph ...\n"
ini_df = df
G = self.construct_graph_from_df(ini_df)
path = save_path + '0/'
os.makedirs(path, exist_ok=True)
grap_mins, graph_message = self.networkx_to_dgl_graph(G, save_path=path)
message += graph_message
print("Initial graph saved")
message += "Initial graph saved\n"
data_split.append(ini_df.shape[0])
all_graph_mins.append(grap_mins)
y = ini_df['event_id'].values
y = [int(each) for each in y]
np.save(path + 'labels.npy', np.asarray(y))
print("Labels saved.")
message += "Labels saved.\n"
indices = ini_df['index'].values.tolist()
x = features[indices, :]
np.save(path + 'features.npy', x)
print("Features saved.")
message += "Features saved.\n\n"
for i in range(7, len(distinct_dates) - 1):
print("Start constructing graph ", str(i - 6), " ...")
message += "\nStart constructing graph "
message += str(i - 6)
message += " ...\n"
incr_df = df.loc[df['date'] == distinct_dates[i]]
if test:
incr_df = incr_df[:test_incr_size]
G = self.construct_graph_from_df(incr_df)
path = save_path + str(i - 6) + '/'
os.makedirs(path, exist_ok=True)
grap_mins, graph_message = self.networkx_to_dgl_graph(G, save_path=path)
message += graph_message
print("Graph ", str(i - 6), " saved")
message += "Graph "
message += str(i - 6)
message += " saved\n"
all_graph_mins.append(grap_mins)
y = [int(each) for each in incr_df['event_id'].values]
np.save(path + 'labels.npy', y)
print("Labels saved.")
message += "Labels saved.\n"
indices = incr_df['index'].values.tolist()
x = features[indices, :]
np.save(path + 'features.npy', x)
print("Features saved.")
message += "Features saved.\n"
return message, data_split, all_graph_mins
[docs] def construct_graph(self, df, save_path='../model/model_saved/finevent/incremental_test/'):
os.makedirs(save_path, exist_ok=True)
df = df.sort_values(by='created_at').reset_index()
df['date'] = [d.date() for d in df['created_at']]
f = np.load('../model/model_saved/finevent/features_69612_0709_spacy_lg_zero_multiclasses_filtered.npy')
message, data_split, all_graph_mins = self.construct_incremental_dataset(df, save_path, f, True)
with open(save_path + "node_edge_statistics.txt", "w") as text_file:
text_file.write(message)
np.save(save_path + 'data_split.npy', np.asarray(data_split))
np.save(save_path + 'all_graph_mins.npy', np.asarray(all_graph_mins))
print("Time spent on heterogeneous -> homogeneous graph conversions: ", all_graph_mins)
[docs] def networkx_to_dgl_graph(self, G, save_path=None):
message = ''
print('Start converting heterogeneous networkx graph to homogeneous dgl graph.')
message += 'Start converting heterogeneous networkx graph to homogeneous dgl graph.\n'
all_start = time()
print('\tGetting a list of all nodes ...')
message += '\tGetting a list of all nodes ...\n'
start = time()
all_nodes = list(G.nodes)
mins = (time() - start) / 60
print('\tDone. Time elapsed: ', mins, ' mins\n')
message += '\tDone. Time elapsed: '
message += str(mins)
message += ' mins\n'
print('\tGetting adjacency matrix ...')
message += '\tGetting adjacency matrix ...\n'
start = time()
A = nx.to_numpy_array(G) # 使用稀疏矩阵
mins = (time() - start) / 60
print('\tDone. Time elapsed: ', mins, ' mins\n')
message += '\tDone. Time elapsed: '
message += str(mins)
message += ' mins\n'
print('\tGetting lists of nodes of various types ...')
message += '\tGetting lists of nodes of various types ...\n'
start = time()
tid_nodes = list(nx.get_node_attributes(G, 'tweet_id').keys())
userid_nodes = list(nx.get_node_attributes(G, 'user_id').keys())
word_nodes = list(nx.get_node_attributes(G, 'word').keys())
entity_nodes = list(nx.get_node_attributes(G, 'entity').keys())
del G
mins = (time() - start) / 60
print('\tDone. Time elapsed: ', mins, ' mins\n')
message += '\tDone. Time elapsed: '
message += str(mins)
message += ' mins\n'
print('\tConverting node lists to index lists ...')
message += '\tConverting node lists to index lists ...\n'
start = time()
indices_tid = [all_nodes.index(x) for x in tid_nodes]
indices_userid = [all_nodes.index(x) for x in userid_nodes]
indices_word = [all_nodes.index(x) for x in word_nodes]
indices_entity = [all_nodes.index(x) for x in entity_nodes]
del tid_nodes
del userid_nodes
del word_nodes
del entity_nodes
mins = (time() - start) / 60
print('\tDone. Time elapsed: ', mins, ' mins\n')
message += '\tDone. Time elapsed: '
message += str(mins)
message += ' mins\n'
# ----------------------tweet-user-tweet----------------------
print('\tStart constructing tweet-user-tweet commuting matrix ...')
print('\t\t\tStart constructing tweet-user matrix ...')
message += '\tStart constructing tweet-user-tweet commuting matrix ...\n\t\t\tStart constructing tweet-user matrix ...\n'
start = time()
w_tid_userid = A[indices_tid, :][:, indices_userid]
mins = (time() - start) / 60
print('\t\t\tDone. Time elapsed: ', mins, ' mins\n')
message += '\t\t\tDone. Time elapsed: '
message += str(mins)
message += ' mins\n'
# convert to scipy sparse matrix
print('\t\t\tConverting to sparse matrix ...')
message += '\t\t\tConverting to sparse matrix ...\n'
start = time()
s_w_tid_userid = csr_matrix(w_tid_userid) # matrix compression
del w_tid_userid
mins = (time() - start) / 60
print('\t\t\tDone. Time elapsed: ', mins, ' mins\n')
message += '\t\t\tDone. Time elapsed: '
message += str(mins)
message += ' mins\n'
print('\t\t\tTransposing ...')
message += '\t\t\tTransposing ...\n'
start = time()
s_w_userid_tid = s_w_tid_userid.transpose()
mins = (time() - start) / 60
print('\t\t\tDone. Time elapsed: ', mins, ' mins\n')
message += '\t\t\tDone. Time elapsed: '
message += str(mins)
message += ' mins\n'
print('\t\t\tCalculating tweet-user * user-tweet ...')
message += '\t\t\tCalculating tweet-user * user-tweet ...\n'
start = time()
s_m_tid_userid_tid = s_w_tid_userid * s_w_userid_tid # homogeneous message graph
mins = (time() - start) / 60
print('\t\t\tDone. Time elapsed: ', mins, ' mins\n')
message += '\t\t\tDone. Time elapsed: '
message += str(mins)
message += ' mins\n'
print('\t\t\tSaving ...')
message += '\t\t\tSaving ...\n'
start = time()
if save_path is not None:
sparse.save_npz(save_path + "s_m_tid_userid_tid.npz", s_m_tid_userid_tid)
print("Sparse binary userid commuting matrix saved.")
del s_m_tid_userid_tid
del s_w_tid_userid
del s_w_userid_tid
mins = (time() - start) / 60
print('\t\t\tDone. Time elapsed: ', mins, ' mins\n')
message += '\t\t\tDone. Time elapsed: '
message += str(mins)
message += ' mins\n'
# ----------------------tweet-ent-tweet------------------------
print('\tStart constructing tweet-ent-tweet commuting matrix ...')
print('\t\t\tStart constructing tweet-ent matrix ...')
message += '\tStart constructing tweet-ent-tweet commuting matrix ...\n\t\t\tStart constructing tweet-ent matrix ...\n'
start = time()
w_tid_entity = A[indices_tid, :][:, indices_entity]
mins = (time() - start) / 60
print('\t\t\tDone. Time elapsed: ', mins, ' mins\n')
message += '\t\t\tDone. Time elapsed: '
message += str(mins)
message += ' mins\n'
# convert to scipy sparse matrix
print('\t\t\tConverting to sparse matrix ...')
message += '\t\t\tConverting to sparse matrix ...\n'
start = time()
s_w_tid_entity = csr_matrix(w_tid_entity)
del w_tid_entity
mins = (time() - start) / 60
print('\t\t\tDone. Time elapsed: ', mins, ' mins\n')
message += '\t\t\tDone. Time elapsed: '
message += str(mins)
message += ' mins\n'
print('\t\t\tTransposing ...')
message += '\t\t\tTransposing ...\n'
start = time()
s_w_entity_tid = s_w_tid_entity.transpose()
mins = (time() - start) / 60
print('\t\t\tDone. Time elapsed: ', mins, ' mins\n')
message += '\t\t\tDone. Time elapsed: '
message += str(mins)
message += ' mins\n'
print('\t\t\tCalculating tweet-ent * ent-tweet ...')
message += '\t\t\tCalculating tweet-ent * ent-tweet ...\n'
start = time()
s_m_tid_entity_tid = s_w_tid_entity * s_w_entity_tid
mins = (time() - start) / 60
print('\t\t\tDone. Time elapsed: ', mins, ' mins\n')
message += '\t\t\tDone. Time elapsed: '
message += str(mins)
message += ' mins\n'
print('\t\t\tSaving ...')
message += '\t\t\tSaving ...\n'
start = time()
if save_path is not None:
sparse.save_npz(save_path + "s_m_tid_entity_tid.npz", s_m_tid_entity_tid)
print("Sparse binary entity commuting matrix saved.")
del s_m_tid_entity_tid
del s_w_tid_entity
del s_w_entity_tid
mins = (time() - start) / 60
print('\t\t\tDone. Time elapsed: ', mins, ' mins\n')
message += '\t\t\tDone. Time elapsed: '
message += str(mins)
message += ' mins\n'
# ----------------------tweet-word-tweet----------------------
print('\tStart constructing tweet-word-tweet commuting matrix ...')
print('\t\t\tStart constructing tweet-word matrix ...')
message += '\tStart constructing tweet-word-tweet commuting matrix ...\n\t\t\tStart constructing tweet-word matrix ...\n'
start = time()
w_tid_word = A[indices_tid, :][:, indices_word]
del A
mins = (time() - start) / 60
print('\t\t\tDone. Time elapsed: ', mins, ' mins\n')
message += '\t\t\tDone. Time elapsed: '
message += str(mins)
message += ' mins\n'
# convert to scipy sparse matrix
print('\t\t\tConverting to sparse matrix ...')
message += '\t\t\tConverting to sparse matrix ...\n'
start = time()
s_w_tid_word = csr_matrix(w_tid_word)
del w_tid_word
mins = (time() - start) / 60
print('\t\t\tDone. Time elapsed: ', mins, ' mins\n')
message += '\t\t\tDone. Time elapsed: '
message += str(mins)
message += ' mins\n'
print('\t\t\tTransposing ...')
message += '\t\t\tTransposing ...\n'
start = time()
s_w_word_tid = s_w_tid_word.transpose()
mins = (time() - start) / 60
print('\t\t\tDone. Time elapsed: ', mins, ' mins\n')
message += '\t\t\tDone. Time elapsed: '
message += str(mins)
message += ' mins\n'
print('\t\t\tCalculating tweet-word * word-tweet ...')
message += '\t\t\tCalculating tweet-word * word-tweet ...\n'
start = time()
s_m_tid_word_tid = s_w_tid_word * s_w_word_tid
mins = (time() - start) / 60
print('\t\t\tDone. Time elapsed: ', mins, ' mins\n')
message += '\t\t\tDone. Time elapsed: '
message += str(mins)
message += ' mins\n'
print('\t\t\tSaving ...')
message += '\t\t\tSaving ...\n'
start = time()
if save_path is not None:
sparse.save_npz(save_path + "s_m_tid_word_tid.npz", s_m_tid_word_tid)
print("Sparse binary word commuting matrix saved.")
del s_m_tid_word_tid
del s_w_tid_word
del s_w_word_tid
mins = (time() - start) / 60
print('\t\t\tDone. Time elapsed: ', mins, ' mins\n')
message += '\t\t\tDone. Time elapsed: '
message += str(mins)
message += ' mins\n'
# ----------------------compute tweet-tweet adjacency matrix----------------------
print('\tComputing tweet-tweet adjacency matrix ...')
message += '\tComputing tweet-tweet adjacency matrix ...\n'
start = time()
if save_path is not None:
s_m_tid_userid_tid = sparse.load_npz(save_path + "s_m_tid_userid_tid.npz")
print("Sparse binary userid commuting matrix loaded.")
s_m_tid_entity_tid = sparse.load_npz(save_path + "s_m_tid_entity_tid.npz")
print("Sparse binary entity commuting matrix loaded.")
s_m_tid_word_tid = sparse.load_npz(save_path + "s_m_tid_word_tid.npz")
print("Sparse binary word commuting matrix loaded.")
s_A_tid_tid = s_m_tid_userid_tid + s_m_tid_entity_tid
del s_m_tid_userid_tid
del s_m_tid_entity_tid
s_bool_A_tid_tid = (s_A_tid_tid + s_m_tid_word_tid).astype(bool) # confirm the connect between tweets
del s_m_tid_word_tid
del s_A_tid_tid
mins = (time() - start) / 60
print('\t\t\tDone. Time elapsed: ', mins, ' mins\n')
message += '\t\t\tDone. Time elapsed: '
message += str(mins)
message += ' mins\n'
all_mins = (time() - all_start) / 60
print('\tOver all time elapsed: ', all_mins, ' mins\n')
message += '\tOver all time elapsed: '
message += str(all_mins)
message += ' mins\n'
if save_path is not None:
sparse.save_npz(save_path + "s_bool_A_tid_tid.npz", s_bool_A_tid_tid)
print("Sparse binary adjacency matrix saved.")
s_bool_A_tid_tid = sparse.load_npz(save_path + "s_bool_A_tid_tid.npz")
print("Sparse binary adjacency matrix loaded.")
# create corresponding dgl graph
G = dgl.from_scipy(s_bool_A_tid_tid)
print('We have %d nodes.' % G.number_of_nodes())
print('We have %d edges.' % G.number_of_edges())
print()
message += 'We have '
message += str(G.number_of_nodes())
message += ' nodes.'
message += 'We have '
message += str(G.number_of_edges())
message += ' edges.\n'
return all_mins, message
[docs] def save_edge_index(self, data_path='../model/model_saved/finevent/incremental_test'):
relation_ids = ['entity', 'userid', 'word']
for i in range(22):
save_multi_relational_graph(data_path, relation_ids, [0, i])
print('edge index saved')
print('all edge index saved')
# gen_dataset
[docs]def sparse_trans(datapath='incremental_test/0/s_m_tid_userid_tid.npz'):
relation = sparse.load_npz(datapath)
all_edge_index = torch.tensor([], dtype=int)
for node in range(relation.shape[0]):
neighbor = torch.IntTensor(relation[node].toarray()).squeeze()
# del self_loop in advance
neighbor[node] = 0
neighbor_idx = neighbor.nonzero()
neighbor_sum = neighbor_idx.size(0)
loop = torch.tensor(node).repeat(neighbor_sum, 1)
edge_index_i_j = torch.cat((loop, neighbor_idx), dim=1).t()
# edge_index_j_i = torch.cat((neighbor_idx, loop), dim=1).t()
self_loop = torch.tensor([[node], [node]])
all_edge_index = torch.cat((all_edge_index, edge_index_i_j, self_loop), dim=1)
del neighbor, neighbor_idx, loop, self_loop, edge_index_i_j
return all_edge_index
[docs]def coo_trans(datapath='incremental_test/0/s_m_tid_userid_tid.npz'):
relation: csr_matrix = sparse.load_npz(datapath)
relation: coo_matrix = relation.tocoo()
sparse_edge_index = torch.LongTensor([relation.row, relation.col])
return sparse_edge_index
[docs]def create_dataset(loadpath, relation, mode):
features = np.load(os.path.join(loadpath, str(mode[1]), 'features.npy'))
features = torch.FloatTensor(features)
print('features loaded')
labels = np.load(os.path.join(loadpath, str(mode[1]), 'labels.npy'))
print('labels loaded')
labels = torch.LongTensor(labels)
relation_edge_index = coo_trans(os.path.join(loadpath, str(mode[1]), 's_m_tid_%s_tid.npz' % relation))
print('edge index loaded')
data = Data(x=features, edge_index=relation_edge_index, y=labels)
data_split = np.load(os.path.join(loadpath, 'data_split.npy'))
train_i, i = mode[0], mode[1]
if train_i == i:
data.train_mask, data.val_mask = generateMasks(len(labels), data_split, train_i, i)
else:
data.test_mask = generateMasks(len(labels), data_split, train_i, i)
return data
[docs]def create_homodataset(loadpath, mode, valid_percent=0.2):
features = np.load(os.path.join(loadpath, str(mode[1]), 'features.npy'))
features = torch.FloatTensor(features)
print('features loaded')
labels = np.load(os.path.join(loadpath, str(mode[1]), 'labels.npy'))
print('labels loaded')
labels = torch.LongTensor(labels)
data = Data(x=features, edge_index=None, y=labels)
data_split = np.load(os.path.join(loadpath, 'data_split.npy'))
train_i, i = mode[0], mode[1]
if train_i == i:
data.train_mask, data.val_mask = generateMasks(len(labels), data_split, train_i, i, valid_percent)
else:
data.test_mask = generateMasks(len(labels), data_split, train_i, i)
return data
[docs]def create_offline_homodataset(loadpath, mode):
features = np.load(os.path.join(loadpath, str(mode[1]), 'features.npy'))
features = torch.FloatTensor(features)
print('features loaded')
labels = np.load(os.path.join(loadpath, str(mode[1]), 'labels.npy'))
print('labels loaded')
labels = torch.LongTensor(labels)
# relation_edge_index = sparse_trans(os.path.join(loadpath, str(mode[1]), 's_bool_A_tid_tid.npz'))
# print('edge index loaded')
data = Data(x=features, edge_index=None, y=labels)
data.train_mask, data.val_mask, data.test_mask = gen_offline_masks(len(labels))
return data
[docs]def create_multi_relational_graph(loadpath, relations, mode):
# multi_relation_edge_index = [sparse_trans(os.path.join(loadpath, str(mode[1]), 's_m_tid_%s_tid.npz' % relation)) for relation in relations]
multi_relation_edge_index = [torch.load(loadpath + '/' + str(mode[1]) + '/edge_index_%s.pt' % relation) for relation
in relations]
print('sparse trans...')
print('edge index loaded')
return multi_relation_edge_index
[docs]def save_multi_relational_graph(loadpath, relations, mode):
for relation in relations:
relation_edge_index = sparse_trans(os.path.join(loadpath, str(mode[1]), 's_m_tid_%s_tid.npz' % relation))
print('%s have saved' % (os.path.join(loadpath, str(mode[1]), 's_m_tid_%s_tid.npz' % relation)))
torch.save(relation_edge_index, loadpath + '/' + str(mode[1]) + '/edge_index_%s.pt' % relation)
# utils
[docs]def intersection(lst1, lst2):
lst3 = [value for value in lst1 if value in lst2]
return lst3
[docs]def run_hdbscan(extract_features, extract_labels, indices, is_validation, isoPath=None, ):
# 2018:min_cluster_size = 5, copy = True, alpha = 0.8
# 2012:min_cluster_size = 8
indices = indices.cpu().detach().numpy()
if isoPath is not None:
# Remove isolated points
temp = torch.load(isoPath)
temp = temp.cpu().detach().numpy()
non_isolated_index = list(np.where(temp != 1)[0])
indices = intersection(indices, non_isolated_index)
# Extract labels
extract_labels = extract_labels.cpu().numpy()
labels_true = extract_labels[indices]
# Extract features
# X = extract_features[indices, :]
X = extract_features.cpu().detach().numpy()
assert labels_true.shape[0] == X.shape[0]
nmi, ami, ari = 0, 0, 0
for eps in [0.2, 0.3, 0.5, 0.7, 1, 1.2, 1.5, 1.7, 2, 2.2, 2.5, 2.7, 3, 3.2, 3.5, 3.7, 4, 4.2, 4.5, 4.7, 5]:
hdb = DBSCAN(eps=eps, min_samples=8)
hdb.fit(X)
labels = hdb.labels_
_nmi = metrics.normalized_mutual_info_score(labels_true, labels)
_ami = metrics.adjusted_mutual_info_score(labels_true, labels)
_ari = metrics.adjusted_rand_score(labels_true, labels)
print(f"_nmi:{_nmi}\t _ami:{_ami}\t _ari:{_ari}\n")
if _nmi > nmi:
nmi = _nmi
ami = _ami
ari = _ari
return nmi, ami, ari
[docs]def run_kmeans(extract_features, extract_labels, indices, isoPath=None):
# Extract the features and labels of the test tweets
indices = indices.cpu().detach().numpy()
if isoPath is not None:
# Remove isolated points
temp = torch.load(isoPath)
temp = temp.cpu().detach().numpy()
non_isolated_index = list(np.where(temp != 1)[0])
indices = intersection(indices, non_isolated_index)
# Extract labels
extract_labels = extract_labels.cpu().numpy()
labels_true = extract_labels[indices]
# Extract features
# X = extract_features[indices, :]
X = extract_features.cpu().detach().numpy()
assert labels_true.shape[0] == X.shape[0]
n_test_tweets = X.shape[0] # 100
# Get the total number of classes
n_classes = len(set(labels_true.tolist()))
# k-means clustering
kmeans = KMeans(n_clusters=n_classes, random_state=0).fit(X)
labels = kmeans.labels_
nmi = metrics.normalized_mutual_info_score(labels_true, labels)
ami = metrics.adjusted_mutual_info_score(labels_true, labels)
ari = metrics.adjusted_rand_score(labels_true, labels)
# Return number of test tweets, number of classes covered by the test tweets, and kMeans cluatering NMI
return n_test_tweets, n_classes, nmi, ami, ari
[docs]def evaluate_model(extract_features, extract_labels, indices,
epoch, num_isolated_nodes, save_path, is_validation=True,
cluster_type='kmeans'):
message = ''
message += '\nEpoch '
message += str(epoch)
message += '\n'
# with isolated nodes
if cluster_type == 'kmeans':
n_tweets, n_classes, nmi, ami, ari = run_kmeans(extract_features, extract_labels, indices)
elif cluster_type == 'dbscan':
pass
if is_validation:
mode = 'validation'
else:
mode = 'test'
message += '\tNumber of ' + mode + ' tweets: '
message += str(n_tweets)
message += '\n\tNumber of classes covered by ' + mode + ' tweets: '
message += str(n_classes)
message += '\n\t' + mode + ' NMI: '
message += str(nmi)
message += '\n\t' + mode + ' AMI: '
message += str(ami)
message += '\n\t' + mode + ' ARI: '
message += str(ari)
if cluster_type == 'dbscan':
message += '\n\t' + mode + ' best_eps: '
# message += str(best_eps)
message += '\n\t' + mode + ' best_min_Pts: '
# message += str(best_min_Pts)
if num_isolated_nodes != 0:
# without isolated nodes
message += '\n\tWithout isolated nodes:'
n_tweets, n_classes, nmi, ami, ari = run_kmeans(extract_features, extract_labels, indices,
save_path + '/isolated_nodes.pt')
message += '\tNumber of ' + mode + ' tweets: '
message += str(n_tweets)
message += '\n\tNumber of classes covered by ' + mode + ' tweets: '
message += str(n_classes)
message += '\n\t' + mode + ' NMI: '
message += str(nmi)
message += '\n\t' + mode + ' AMI: '
message += str(ami)
message += '\n\t' + mode + ' ARI: '
message += str(ari)
message += '\n'
with open(save_path + '/evaluate.txt', 'a') as f:
f.write(message)
print(message)
np.save(save_path + '/%s_metric.npy' % mode, np.asarray([nmi, ami, ari]))
return nmi
[docs]def generateMasks(length, data_split, train_i, i, validation_percent=0.2, save_path=None, remove_obsolete=2):
"""
Intro:
This function generates train and validation indices for initial/maintenance epochs and test indices for inference(prediction) epochs
If remove_obsolete mode 0 or 1:
For initial/maintenance epochs:
- The first (train_i + 1) blocks (blocks 0, ..., train_i) are used as training set (with explicit labels)
- Randomly sample validation_percent of the training indices as validation indices
For inference(prediction) epochs:
- The (i + 1)th block (block i) is used as test set
Note that other blocks (block train_i + 1, ..., i - 1) are also in the graph (without explicit labels, only their features and structural info are leveraged)
If remove_obsolete mode 2:
For initial/maintenance epochs:
- The (i + 1) = (train_i + 1)th block (block train_i = i) is used as training set (with explicit labels)
- Randomly sample validation_percent of the training indices as validation indices
For inference(prediction) epochs:
- The (i + 1)th block (block i) is used as test set
:param length: the length of label list
:param data_split: loaded splited data (generated in custom_message_graph.py)
:param train_i, i: flag, indicating for initial/maintenance stage if train_i == i and inference stage for others
:param validation_percent: the percent of validation data occupied in whole dataset
:param save_path: path to save data
:param num_indices_to_remove: number of indices ought to be removed
:returns train indices, validation indices or test indices
"""
print(length)
print(data_split[i])
# step1: verify total number of nodes
assert length == data_split[i] # 500
# step2.0: if is in initial/maintenance epochs, generate train and validation indices
if train_i == i:
# step3: randomly shuffle the graph indices
train_indices = torch.randperm(length)
# step4: get total number of validation indices
n_validation_samples = int(length * validation_percent)
# step5: sample n_validation_samples validation indices and use the rest as training indices
validation_indices = train_indices[:n_validation_samples]
train_indices = train_indices[n_validation_samples:]
# print(save_path) #./incremental_0808//embeddings_0403100832/block_0/masks
# step6: save indices
if save_path is not None:
torch.save(train_indices, save_path + '/train_indices.pt')
torch.save(validation_indices, save_path + '/validation_indices.pt')
return train_indices, validation_indices
# step2.1: if is in inference(prediction) epochs, generate test indices
else:
test_indices = torch.arange(0, (data_split[i]), dtype=torch.long)
if save_path is not None:
torch.save(test_indices, save_path + '/test_indices.pt')
return test_indices
[docs]def gen_offline_masks(length, validation_percent=0.2, test_percent=0.1):
test_length = int(length * test_percent)
valid_length = int(length * validation_percent)
train_length = length - valid_length - test_length
samples = torch.randperm(length)
train_indices = samples[:train_length]
valid_indices = samples[train_length:train_length + valid_length]
test_indices = samples[train_length + valid_length:]
return train_indices, valid_indices, test_indices
[docs]def save_embeddings(extracted_features, save_path):
torch.save(extracted_features, save_path + '/final_embeddings.pt')
print('extracted features saved.')
# Mysampler
[docs]class MySampler(object):
def __init__(self, sampler) -> None:
super().__init__()
self.sampler = sampler
[docs] def sample(self, multi_relational_edge_index: List[Tensor], node_idx, sizes, batch_size):
if self.sampler == 'RL_sampler':
return self._RL_sample(multi_relational_edge_index, node_idx, sizes, batch_size)
elif self.sampler == 'random_sampler':
return self._random_sample(multi_relational_edge_index, node_idx, batch_size)
elif self.sampler == 'const_sampler':
return self._const_sample(multi_relational_edge_index, node_idx, batch_size)
def _RL_sample(self, multi_relational_edge_index: List[Tensor], node_idx, sizes, batch_size):
outs = []
all_n_ids = []
for id, edge_index in enumerate(multi_relational_edge_index):
loader = NeighborSampler(edge_index=edge_index,
sizes=sizes,
node_idx=node_idx,
return_e_id=False,
batch_size=batch_size,
num_workers=0)
for id, (_, n_ids, adjs) in enumerate(loader):
# print(adjs)
outs.append(adjs)
all_n_ids.append(n_ids)
# print(id)
assert id == 0
return outs, all_n_ids
def _random_sample(self, multi_relational_edge_index: List[Tensor], node_idx, batch_size):
outs = []
all_n_ids = []
sizes = [random.randint(10, 100), random.randint(10, 50)]
for edge_index in multi_relational_edge_index:
loader = NeighborSampler(edge_index=edge_index,
sizes=sizes,
node_idx=node_idx,
return_e_id=False,
batch_size=batch_size,
num_workers=0)
for id, (_, n_ids, adjs) in enumerate(loader):
# print(adjs)
outs.append(adjs)
all_n_ids.append(n_ids)
# print(id)
assert id == 0
return outs, all_n_ids
def _const_sample(self, multi_relational_edge_index: List[Tensor], node_idx, batch_size):
outs = []
all_n_ids = []
sizes = [25, 15]
for edge_index in multi_relational_edge_index:
loader = NeighborSampler(edge_index=edge_index,
sizes=sizes,
node_idx=node_idx,
return_e_id=False,
batch_size=batch_size,
num_workers=0)
for id, (_, n_ids, adjs) in enumerate(loader):
# print(adjs)
outs.append(adjs)
all_n_ids.append(n_ids)
# print(id)
assert id == 0
return outs, all_n_ids
# Metrics
[docs]class Metric:
def __init__(self):
pass
def __call__(self, outputs, target, loss):
raise NotImplementedError
[docs] def reset(self):
raise NotImplementedError
[docs] def value(self):
raise NotImplementedError
[docs] def name(self):
raise NotImplementedError
[docs]class AccumulatedAccuracyMetric(Metric):
"""
Works with classification model
"""
def __init__(self):
self.correct = 0
self.total = 0
def __call__(self, outputs, target, loss):
pred = outputs[0].data.max(1, keepdim=True)[1]
self.correct += pred.eq(target[0].data.view_as(pred)).cpu().sum()
self.total += target[0].size(0)
return self.value()
[docs] def reset(self):
self.correct = 0
self.total = 0
[docs] def value(self):
return 100 * float(self.correct) / self.total
[docs] def name(self):
return 'Accuracy'
[docs]class AverageNonzeroTripletsMetric(Metric):
'''
Counts average number of nonzero triplets found in minibatches
'''
def __init__(self):
self.values = []
def __call__(self, outputs, target, loss):
self.values.append(loss[1])
return self.value()
[docs] def reset(self):
self.values = []
[docs] def value(self):
return np.mean(self.values)
[docs] def name(self):
return 'Average nonzero triplets'
# model
[docs]class MarGNN(nn.Module):
def __init__(self, GNN_args, num_relations, inter_opt, is_shared=False):
super(MarGNN, self).__init__()
self.num_relations = num_relations
self.inter_opt = inter_opt
self.is_shared = is_shared
if not self.is_shared:
self.intra_aggs = torch.nn.ModuleList([Intra_AGG(GNN_args) for _ in range(self.num_relations)])
else:
self.intra_aggs = Intra_AGG(GNN_args) # shared parameters
if self.inter_opt == 'cat_w_avg_mlp' or 'cat_wo_avg_mlp':
in_dim, hid_dim, out_dim, heads = GNN_args
mlp_args = self.num_relations * out_dim, out_dim
self.inter_agg = Inter_AGG(mlp_args)
else:
self.inter_agg = Inter_AGG()
[docs] def forward(self, x, adjs, n_ids, device, RL_thresholds):
# RL_threshold: tensor([[.5], [.5], [.5]])
if RL_thresholds is None:
RL_thresholds = torch.FloatTensor([[1.], [1.], [1.]])
if not isinstance(RL_thresholds, Tensor):
RL_thresholds = torch.FloatTensor(RL_thresholds)
RL_thresholds = RL_thresholds.to(device)
features = []
for i in range(self.num_relations):
if not self.is_shared:
# print('Intra Aggregation of relation %d' % i)
features.append(self.intra_aggs[i](x[n_ids[i]], adjs[i], device))
else:
# shared parameters.
# print('Shared Intra Aggregation...')
features.append(self.intra_aggs(x[n_ids[i]], adjs[i], device))
features = torch.stack(features, dim=0)
features = self.inter_agg(features, RL_thresholds, self.inter_opt)
return features
# env
[docs]def RL_neighbor_filter_full(multi_r_data, RL_thresholds, features, save_path=None):
multi_remain_data = []
multi_r_score = []
for i, r_data in enumerate(multi_r_data):
r_data: Tensor
unique_nodes = r_data[1].unique()
num_nodes = unique_nodes.size(0)
remain_node_index = torch.tensor([])
node_scores = []
for node in range(num_nodes):
# get neighbors' index
neighbors_idx = torch.where(r_data[1] == node)[0]
# get neighbors
neighbors = r_data[0, neighbors_idx]
num_neighbors = neighbors.size(0)
neighbors_features = features[neighbors, :]
target_features = features[node, :]
# calculate euclid distance with broadcast
dist: Tensor = torch.norm(neighbors_features - target_features, p=2, dim=1)
# smaller is better and we use 'top p' in our paper
# => (threshold * num_neighbors)
# see RL_neighbor_filter for details
sorted_neighbors, sorted_index = dist.sort(descending=False)
if num_neighbors <= 5:
remain_node_index = torch.cat((remain_node_index, neighbors_idx))
continue # add limitations
threshold = float(RL_thresholds[i])
num_kept_neighbors = math.ceil(num_neighbors * threshold) + 1
filtered_neighbors_idx = neighbors_idx[sorted_index[:num_kept_neighbors]]
remain_node_index = torch.cat((remain_node_index, filtered_neighbors_idx))
filtered_neighbors_scores = sorted_neighbors[:num_kept_neighbors].mean()
node_scores.append(filtered_neighbors_scores)
remain_node_index = remain_node_index.type('torch.LongTensor')
edge_index = r_data[:, remain_node_index]
multi_remain_data.append(edge_index)
node_scores = torch.FloatTensor(node_scores) # from list
avg_node_scores = node_scores.sum(dim=1) / num_nodes
multi_r_score.append(avg_node_scores)
return multi_remain_data, multi_r_score
[docs]def multi_forward_agg(args, foward_args, iter_epoch):
# args prepare
model, homo_data, all_num_samples, num_dim, sampler, multi_r_data, filtered_multi_r_data, device, RL_thresholds = foward_args
if filtered_multi_r_data is None:
filtered_multi_r_data = multi_r_data
extract_features = torch.FloatTensor([])
num_batches = int(all_num_samples / args.batch_size) + 1
# all mask are then splited into mini-batch in order
all_mask = torch.arange(0, num_dim, dtype=torch.long)
# multiple forward with RL training
for _ in range(iter_epoch):
# batch training
for batch in range(num_batches):
start_batch = time()
# split batch
i_start = args.batch_size * batch
i_end = min((batch + 1) * args.batch_size, all_num_samples)
batch_nodes = all_mask[i_start:i_end]
batch_labels = homo_data.y[batch_nodes]
# sampling neighbors of batch nodes
adjs, n_ids = sampler.sample(filtered_multi_r_data, node_idx=batch_nodes, sizes=[-1, -1],
batch_size=args.batch_size)
pred = model(homo_data.x, adjs, n_ids, device, RL_thresholds)
extract_features = torch.cat((extract_features, pred.cpu().detach()), dim=0)
del pred
# RL trainig
filtered_multi_r_data, multi_r_scores = RL_neighbor_filter_full(filtered_multi_r_data, RL_thresholds,
extract_features)
# return new RL thresholds
return RL_thresholds
# layer
[docs]class GAT(nn.Module):
'''
adopt this module when using mini-batch
'''
def __init__(self, in_dim, hid_dim, out_dim, heads) -> None:
super(GAT, self).__init__()
self.GAT1 = GATConv(in_channels=in_dim, out_channels=hid_dim, heads=heads, add_self_loops=False)
self.GAT2 = GATConv(in_channels=hid_dim * heads, out_channels=out_dim, add_self_loops=False)
self.layers = ModuleList([self.GAT1, self.GAT2])
self.norm = BatchNorm1d(heads * hid_dim)
[docs] def forward(self, x, adjs, device):
for i, (edge_index, _, size) in enumerate(adjs):
# x: Tensor, edge_index: Tensor
x, edge_index = x.to(device), edge_index.to(device)
x_target = x[:size[1]] # Target nodes are always placed first.
x = self.layers[i]((x, x_target), edge_index)
if i == 0:
x = self.norm(x)
x = F.elu(x)
x = F.dropout(x, training=self.training)
del edge_index
return x
[docs]class Intra_AGG(nn.Module):
def __init__(self, GAT_args):
super(Intra_AGG, self).__init__()
in_dim, hid_dim, out_dim, heads = GAT_args
self.gnn = GAT(in_dim, hid_dim, out_dim, heads)
[docs] def forward(self, x, adjs, device):
x = self.gnn(x, adjs, device)
return x
[docs]class Inter_AGG(nn.Module):
def __init__(self, mlp_args=None):
super(Inter_AGG, self).__init__()
if mlp_args is not None:
hid_dim, out_dim = mlp_args
self.mlp = nn.Sequential(
Linear(hid_dim, hid_dim),
BatchNorm1d(hid_dim),
ReLU(inplace=True),
Dropout(),
Linear(hid_dim, out_dim),
)
[docs] def forward(self, features, thresholds, inter_opt):
batch_size = features[0].size(0)
features = torch.transpose(features, dim0=0, dim1=1)
if inter_opt == 'cat_wo_avg':
features = features.reshape(batch_size, -1)
elif inter_opt == 'cat_w_avg':
# weighted average and concatenate
features = torch.mul(features, thresholds).reshape(batch_size, -1)
elif inter_opt == 'cat_w_avg_mlp':
features = torch.mul(features, thresholds).reshape(batch_size, -1)
features = self.mlp(features)
elif inter_opt == 'cat_wo_avg_mlp':
features = torch.mul(features, thresholds).reshape(batch_size, -1)
features = self.mlp(features)
elif inter_opt == 'add_wo_avg':
features = features.sum(dim=1)
elif inter_opt == 'add_w_avg':
features = torch.mul(features, thresholds).sum(dim=1)
return features
# neighborRL
[docs]def pre_node_dist(multi_r_data, features, save_path=None):
"""This is used to culculate the similarity between node and
its neighbors in advance in order to avoid the repetitive computation.
Args:
multi_r_data ([type]): [description]
features ([type]): [description]
save_path ([type], optional): [description]. Defaults to None.
"""
relation_config: Dict[str, Dict[int, Any]] = {}
for relation_id, r_data in enumerate(multi_r_data):
node_config: Dict[int, Any] = {}
r_data: Tensor
unique_nodes = r_data[1].unique()
num_nodes = unique_nodes.size(0)
for node in range(num_nodes):
# get neighbors' index
neighbors_idx = torch.where(r_data[1] == node)[0]
# get neighbors
neighbors = r_data[0, neighbors_idx]
num_neighbors = neighbors.size(0)
neighbors_features = features[neighbors, :]
target_features = features[node, :]
# calculate euclid distance with broadcast
dist: Tensor = torch.norm(neighbors_features - target_features, p=2, dim=1)
# smaller is better and we use 'top p' in our paper
# (threshold * num_neighbors) see RL_neighbor_filter for details
sorted_neighbors, sorted_index = dist.sort(descending=False)
node_config[node] = {'neighbors_idx': neighbors_idx,
'sorted_neighbors': sorted_neighbors,
'sorted_index': sorted_index,
'num_neighbors': num_neighbors}
relation_config['relation_%d' % relation_id] = node_config
if save_path is not None:
save_path = os.path.join(save_path, 'relation_config.npy')
# print(save_path)
np.save(save_path, relation_config)
[docs]def RL_neighbor_filter(args,multi_r_data, RL_thresholds, load_path):
load_path = os.path.join(load_path, 'relation_config.npy')
relation_config = np.load(load_path, allow_pickle=True)
relation_config = relation_config.tolist()
relations = list(relation_config.keys())
multi_remain_data = []
for i in range(len(relations)):
print(f"Processing relation {i + 1}/{len(relations)}: {relations[i]}")
edge_index: Tensor = multi_r_data[i]
unique_nodes = edge_index[1].unique()
num_nodes = unique_nodes.size(0)
remain_node_index = torch.tensor([])
for node in range(num_nodes):
if node % 1000 == 0: # 每处理1000个节点输出一次进度
print(f" Processing node {node}/{num_nodes}")
# extract config
neighbors_idx = relation_config[relations[i]][node]['neighbors_idx']
num_neighbors = relation_config[relations[i]][node]['num_neighbors']
sorted_neighbors = relation_config[relations[i]][node]['sorted_neighbors']
sorted_index = relation_config[relations[i]][node]['sorted_index']
if num_neighbors <= 5:
remain_node_index = torch.cat((remain_node_index, neighbors_idx))
continue # add limitations
threshold = float(RL_thresholds[i])
num_kept_neighbors = math.ceil(num_neighbors * threshold) + 1
filtered_neighbors_idx = neighbors_idx[sorted_index[:num_kept_neighbors]]
# 修正超出范围的索引
valid_indices = filtered_neighbors_idx[filtered_neighbors_idx < edge_index.size(1)]
remain_node_index = torch.cat((remain_node_index, valid_indices))
remain_node_index = remain_node_index.type('torch.LongTensor')
# print(remain_node_index)
# Debugging print statements
max_index = remain_node_index.max().item()
edge_size = edge_index.size(1)
print(f"Max remain_node_index: {max_index}")
print(f"Edge index size: {edge_size}")
# 修正索引超出范围的情况
if max_index >= edge_size:
remain_node_index = remain_node_index[remain_node_index < edge_size]
edge_index = edge_index[:, remain_node_index]
multi_remain_data.append(edge_index)
print(f"Finished processing relation {relations[i]}")
# 保存 multi_remain_data
save_path = os.path.join(args.file_path, 'multi_remain_data.pt')
torch.save(multi_remain_data, save_path)
print(f"Filtered multi_r_data saved successfully at {save_path}")
return multi_remain_data
# TripletLoss
[docs]class AvgReadout(nn.Module):
def __init__(self):
super(AvgReadout, self).__init__()
[docs] def forward(self, seq):
return torch.mean(seq, 0)
[docs]class Discriminator(nn.Module):
def __init__(self, n_h):
super(Discriminator, self).__init__()
self.f_k = nn.Bilinear(n_h, n_h, 1)
for m in self.modules():
self.weights_init(m)
[docs] def weights_init(self, m):
if isinstance(m, nn.Bilinear):
torch.nn.init.xavier_uniform_(m.weight.data)
if m.bias is not None:
m.bias.data.fill_(0.0)
[docs] def forward(self, c, h_pl, h_mi, s_bias1=None, s_bias2=None):
c_x = torch.unsqueeze(c, 0)
c_x = c_x.expand_as(h_pl)
sc_1 = torch.squeeze(self.f_k(h_pl, c_x), 1)
sc_2 = torch.squeeze(self.f_k(h_mi, c_x), 1)
if s_bias1 is not None:
sc_1 += s_bias1
if s_bias2 is not None:
sc_2 += s_bias2
logits = torch.cat((sc_1, sc_2), 0)
# print("testing, shape of logits: ", logits.size())
return logits
[docs]class OnlineTripletLoss(nn.Module):
"""
Online Triplets loss
Takes a batch of embeddings and corresponding labels.
Triplets are generated using triplet_selector object that take embeddings and targets and return indices of triplets.
"""
def __init__(self, margin, triplet_selector):
super(OnlineTripletLoss, self).__init__()
self.margin = margin
self.triplet_selector = triplet_selector
[docs] def forward(self, embeddings, target):
# 确保 embeddings 至少是二维
if embeddings.dim() == 1:
embeddings = embeddings.unsqueeze(1)
triplets = self.triplet_selector.get_triplets(embeddings, target)
# 打印 embeddings 和 triplets 的形状用于调试
print(f"embeddings shape: {embeddings.shape}")
print(f"triplets shape: {triplets.shape}")
# 检查 triplets 是否为空
if triplets.numel() == 0:
return torch.tensor(0.0, requires_grad=True), 0
# 确保 triplets 的索引在 embeddings 的范围内
if (triplets >= embeddings.size(0)).any():
raise IndexError("triplets index out of range of embeddings")
# 计算 anchor-positive 和 anchor-negative 的距离
ap_distances = (embeddings[triplets[:, 0]] - embeddings[triplets[:, 1]]).pow(2).sum(1) # .pow(.5)
an_distances = (embeddings[triplets[:, 0]] - embeddings[triplets[:, 2]]).pow(2).sum(1) # .pow(.5)
losses = F.relu(ap_distances - an_distances + self.margin)
return losses.mean(), len(triplets)
[docs]def pdist(vectors):
distance_matrix = -2 * vectors.mm(torch.t(vectors)) + vectors.pow(2).sum(dim=1).view(1, -1) + vectors.pow(2).sum(
dim=1).view(-1, 1)
return distance_matrix
[docs]class TripletSelector:
"""
Implementation should return indices of anchors, positive and negative samples
return np array of shape [N_triplets x 3]
"""
def __init__(self):
pass
[docs] def get_triplets(self, embeddings, labels):
raise NotImplementedError
[docs]class FunctionNegativeTripletSelector(TripletSelector):
"""
For each positive pair, takes the hardest negative sample (with the greatest triplet loss value) to create a triplet
Margin should match the margin used in triplet loss.
negative_selection_fn should take array of loss_values for a given anchor-positive pair and all negative samples
and return a negative index for that pair
"""
def __init__(self, margin, negative_selection_fn, cpu=True):
super(FunctionNegativeTripletSelector, self).__init__()
self.cpu = cpu
self.margin = margin
self.negative_selection_fn = negative_selection_fn
[docs] def get_triplets(self, embeddings, labels):
if self.cpu:
embeddings = embeddings.cpu()
distance_matrix = pdist(embeddings)
distance_matrix = distance_matrix.cpu()
labels = labels.cpu().data.numpy()
triplets = []
for label in set(labels):
label_mask = (labels == label)
label_indices = np.where(label_mask)[0]
if len(label_indices) < 2:
continue
negative_indices = np.where(np.logical_not(label_mask))[0]
anchor_positives = list(combinations(label_indices, 2)) # All anchor-positive pairs
anchor_positives = np.array(anchor_positives)
ap_distances = distance_matrix[anchor_positives[:, 0], anchor_positives[:, 1]]
for anchor_positive, ap_distance in zip(anchor_positives, ap_distances):
loss_values = ap_distance - distance_matrix[
torch.LongTensor(np.array([anchor_positive[0]])), torch.LongTensor(negative_indices)] + self.margin
loss_values = loss_values.data.cpu().numpy()
hard_negative = self.negative_selection_fn(loss_values)
if hard_negative is not None:
hard_negative = negative_indices[hard_negative]
triplets.append([anchor_positive[0], anchor_positive[1], hard_negative])
# if len(triplets) == 0:
# triplets.append([anchor_positive[0], anchor_positive[1], negative_indices[0]])
triplets = np.array(triplets)
return torch.LongTensor(triplets)
[docs]def random_hard_negative(loss_values):
hard_negatives = np.where(loss_values > 0)[0]
return np.random.choice(hard_negatives) if len(hard_negatives) > 0 else None
[docs]def hardest_negative(loss_values):
hard_negative = np.argmax(loss_values)
return hard_negative if loss_values[hard_negative] > 0 else None
[docs]def HardestNegativeTripletSelector(margin, cpu=False): return FunctionNegativeTripletSelector(margin=margin,
negative_selection_fn=hardest_negative,
cpu=cpu)
[docs]def RandomNegativeTripletSelector(margin, cpu=False): return FunctionNegativeTripletSelector(margin=margin,
negative_selection_fn=random_hard_negative,
cpu=cpu)