cs3319-project2 / code /run_improved.py
NLP-beginner's picture
CS3319 Project 2 final deliverable (public F1 = 0.96626)
f28d994
Raw
History Blame Contribute Delete
23 kB
"""Improved heterogeneous GNN for academic paper recommendation.
Key improvements over baseline:
1. Proper SAGEConv-based heterogeneous GNN (3 layers, residual)
2. MLP decoder instead of dot product
3. Hard negative sampling (popular papers + co-author papers)
4. Graph structural features (degree features)
5. BCE loss with positive weight
6. Longer training with LR scheduling
7. Exploits train-test overlap for known positives
"""
import os
import pickle as pkl
import random
import itertools
import numpy as np
import pandas as pd
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.optim.lr_scheduler import CosineAnnealingLR, ReduceLROnPlateau
from torch_geometric.data import HeteroData
from torch_geometric.nn import SAGEConv, HeteroConv, Linear
from sklearn.metrics import f1_score, precision_recall_curve, roc_auc_score
device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
print('device:', device)
def set_seed(seed=0):
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
if torch.cuda.is_available():
torch.cuda.manual_seed_all(seed)
# ── Data loading ──────────────────────────────────────────────────
base_path = "/home/lzc/cs3319-project"
def read_txt(file):
res_list = []
with open(file, "r") as f:
for line in f:
res_list.append(list(map(int, line.strip().split())))
return res_list
citation = read_txt(os.path.join(base_path, "paper_file_ann.txt"))
existing_refs = read_txt(os.path.join(base_path, "bipartite_train_ann.txt"))
refs_to_pred = read_txt(os.path.join(base_path, "bipartite_test_ann.txt"))
coauthor = read_txt(os.path.join(base_path, "author_file_ann.txt"))
with open(os.path.join(base_path, "feature.pkl"), 'rb') as f:
paper_feature = pkl.load(f)
print(f"Citations: {len(citation)}, Train refs: {len(existing_refs)}, "
f"Test pairs: {len(refs_to_pred)}, Coauthor: {len(coauthor)}")
# ── Compute train-test overlap ────────────────────────────────────
train_set = set(map(tuple, existing_refs))
test_arr_full = np.array(refs_to_pred, dtype=np.int64)
test_set = set(map(tuple, refs_to_pred))
overlap = train_set & test_set
print(f"Train-test overlap (known positives): {len(overlap)} / {len(test_set)} "
f"({100*len(overlap)/len(test_set):.1f}%)")
# ── Convert to DataFrames ────────────────────────────────────────
cite_edges = pd.DataFrame(citation, columns=['source', 'target'])
ref_edges = pd.DataFrame(existing_refs, columns=['source', 'target'])
coauthor_edges = pd.DataFrame(coauthor, columns=['source', 'target'])
# Build node sets
node_tmp = pd.concat([cite_edges['source'], cite_edges['target'], ref_edges['target']])
node_papers = pd.DataFrame(index=pd.unique(node_tmp))
node_tmp = pd.concat([ref_edges['source'], coauthor_edges['source'], coauthor_edges['target']])
node_authors = pd.DataFrame(index=pd.unique(node_tmp))
num_paper_nodes = len(node_papers)
num_author_nodes = len(node_authors)
print(f"Nodes: {num_author_nodes} authors, {num_paper_nodes} papers")
# ── Degree features ───────────────────────────────────────────────
author_ref_deg = np.zeros(num_author_nodes, dtype=np.float32)
paper_ref_deg = np.zeros(num_paper_nodes, dtype=np.float32)
author_coauthor_deg = np.zeros(num_author_nodes, dtype=np.float32)
paper_cite_out = np.zeros(num_paper_nodes, dtype=np.float32)
paper_cite_in = np.zeros(num_paper_nodes, dtype=np.float32)
for s, t in existing_refs:
author_ref_deg[s] += 1
paper_ref_deg[t] += 1
for s, t in coauthor:
author_coauthor_deg[s] += 1
author_coauthor_deg[t] += 1
for s, t in citation:
paper_cite_out[s] += 1
paper_cite_in[t] += 1
# Log-transform and normalize degree features
def log_norm(x):
x = np.log1p(x)
return (x - x.mean()) / (x.std() + 1e-8)
author_deg_feat = np.stack([
log_norm(author_ref_deg),
log_norm(author_coauthor_deg),
], axis=-1)
paper_deg_feat = np.stack([
log_norm(paper_ref_deg),
log_norm(paper_cite_out),
log_norm(paper_cite_in),
], axis=-1)
# Augment paper features with degree features
paper_feature_np = paper_feature.numpy()
paper_feat_aug = np.concatenate([paper_feature_np, paper_deg_feat], axis=-1)
paper_feat_dim = paper_feat_aug.shape[-1]
author_deg_dim = author_deg_feat.shape[-1]
print(f"Paper features: {paper_feat_dim}d (512 + {paper_deg_feat.shape[-1]} degree), "
f"Author degree features: {author_deg_feat.shape[-1]}d")
# ── Train/val split (90/10) ──────────────────────────────────────
ref_edges_idx = ref_edges.copy()
train_refs = ref_edges_idx.sample(frac=0.9, random_state=0, axis=0)
val_pos = ref_edges_idx[~ref_edges_idx.index.isin(train_refs.index)].copy()
val_pos['label'] = 1
# Validation negatives
existing_ref_set = set(map(tuple, existing_refs))
author_ids = node_authors.index.to_numpy(dtype=np.int64)
paper_ids = node_papers.index.to_numpy(dtype=np.int64)
num_val_neg = len(val_pos)
neg_pairs = []
rng = np.random.default_rng(0)
while len(neg_pairs) < num_val_neg:
src = int(rng.choice(author_ids))
dst = int(rng.choice(paper_ids))
if (src, dst) not in existing_ref_set:
neg_pairs.append((src, dst))
val_neg = pd.DataFrame(neg_pairs, columns=['source', 'target'])
val_neg['label'] = 0
val_set = pd.concat([val_pos, val_neg], ignore_index=True).sample(frac=1, random_state=0)
print(f"Val: {len(val_set)} ({val_set['label'].sum()} pos, {len(val_set)-val_set['label'].sum()} neg)")
# ── Pre-compute hard negative pools ───────────────────────────────
# Popular papers: top papers by reference degree
paper_popularity = paper_ref_deg.copy()
popular_threshold = np.percentile(paper_popularity[paper_popularity > 0], 70)
popular_papers = np.where(paper_popularity >= popular_threshold)[0]
print(f"Popular papers (top 30%): {len(popular_papers)}")
# Co-author mapping
coauthor_map = {i: set() for i in range(num_author_nodes)}
for s, t in coauthor:
coauthor_map[s].add(t)
coauthor_map[t].add(s)
# Author's known papers
author_papers = {i: set() for i in range(num_author_nodes)}
for s, t in existing_refs:
author_papers[s].add(t)
# Co-author's papers (papers read by at least one co-author but not by this author)
coauthor_paper_pool = {}
for author in range(num_author_nodes):
pool = set()
for coa in coauthor_map[author]:
pool.update(author_papers[coa])
pool -= author_papers[author]
coauthor_paper_pool[author] = list(pool) if pool else list(range(num_paper_nodes))
print(f"Authors with co-author paper pool: "
f"{sum(1 for v in coauthor_paper_pool.values() if len(v) > 0)}")
# ── Build HeteroData ──────────────────────────────────────────────
train_ref_tensor = torch.as_tensor(train_refs[['source', 'target']].to_numpy(), dtype=torch.long)
cite_tensor = torch.as_tensor(cite_edges[['source', 'target']].to_numpy(), dtype=torch.long)
coauthor_tensor = torch.as_tensor(coauthor_edges[['source', 'target']].to_numpy(), dtype=torch.long)
num_authors = num_author_nodes
num_papers = num_paper_nodes
paper_x = torch.as_tensor(paper_feat_aug, dtype=torch.float)
author_x = torch.as_tensor(author_deg_feat, dtype=torch.float)
data = HeteroData()
data['author'].num_nodes = num_authors
data['author'].x = author_x
data['paper'].num_nodes = num_papers
data['paper'].x = paper_x
data['author', 'ref', 'paper'].edge_index = train_ref_tensor.t().contiguous()
data['paper', 'beref', 'author'].edge_index = train_ref_tensor[:, [1, 0]].t().contiguous()
data['paper', 'cite', 'paper'].edge_index = torch.cat([
cite_tensor, cite_tensor[:, [1, 0]],
], dim=0).t().contiguous()
data['author', 'coauthor', 'author'].edge_index = torch.cat([
coauthor_tensor, coauthor_tensor[:, [1, 0]],
], dim=0).t().contiguous()
data = data.to(device)
print(data)
# ── Model ─────────────────────────────────────────────────────────
class ResidualHeteroConv(nn.Module):
"""HeteroConv with residual connection and layer norm."""
def __init__(self, metadata, in_dims, out_dim, dropout=0.2):
super().__init__()
node_types, edge_types = metadata
edge_types_used = [
('author', 'ref', 'paper'),
('paper', 'beref', 'author'),
('paper', 'cite', 'paper'),
('author', 'coauthor', 'author'),
]
conv_dict = {}
for et in edge_types_used:
if et in edge_types:
conv_dict[et] = SAGEConv(
(in_dims[et[0]], in_dims[et[2]]), out_dim,
)
self.conv = HeteroConv(conv_dict, aggr='mean')
self.norms = nn.ModuleDict({
nt: nn.LayerNorm(out_dim) for nt in node_types
})
self.dropout = nn.Dropout(dropout)
# Projection for residual
self.res_proj = nn.ModuleDict()
for nt in node_types:
if in_dims.get(nt, out_dim) != out_dim:
self.res_proj[nt] = nn.Linear(in_dims[nt], out_dim)
def forward(self, x_dict, edge_index_dict):
new_x = self.conv(x_dict, edge_index_dict)
out = {}
for nt in new_x:
res = x_dict[nt]
if nt in self.res_proj:
res = self.res_proj[nt](res)
out[nt] = self.norms[nt](new_x[nt] + res)
out[nt] = F.relu(out[nt])
out[nt] = self.dropout(out[nt])
return out
class MLPDecoder(nn.Module):
"""MLP decoder: author_emb || paper_emb || author_emb * paper_emb -> score."""
def __init__(self, in_dim, hidden=128):
super().__init__()
self.mlp = nn.Sequential(
nn.Linear(in_dim * 3, hidden),
nn.BatchNorm1d(hidden),
nn.ReLU(),
nn.Dropout(0.3),
nn.Linear(hidden, hidden // 2),
nn.BatchNorm1d(hidden // 2),
nn.ReLU(),
nn.Dropout(0.3),
nn.Linear(hidden // 2, 1),
)
def forward(self, author_emb, paper_emb, edge_label_index):
src, dst = edge_label_index
a = author_emb[src]
p = paper_emb[dst]
x = torch.cat([a, p, a * p], dim=-1)
return self.mlp(x).squeeze(-1)
class ImprovedHeteroGNN(nn.Module):
def __init__(self, metadata, author_in_dim, paper_in_dim,
hidden_dim=128, num_layers=3, out_dim=64):
super().__init__()
node_types, edge_types = metadata
self.author_proj = nn.Linear(author_in_dim, hidden_dim)
self.paper_proj = nn.Linear(paper_in_dim, hidden_dim)
in_dims_init = {'author': hidden_dim, 'paper': hidden_dim}
self.convs = nn.ModuleList()
for i in range(num_layers):
self.convs.append(ResidualHeteroConv(
metadata, in_dims_init, hidden_dim, dropout=0.2,
))
in_dims_init = {'author': hidden_dim, 'paper': hidden_dim}
self.post_lin = nn.Linear(hidden_dim, out_dim)
self.decoder = MLPDecoder(out_dim, hidden=128)
self.reset_parameters()
def reset_parameters(self):
for m in self.modules():
if isinstance(m, nn.Linear):
nn.init.xavier_uniform_(m.weight)
if m.bias is not None:
nn.init.zeros_(m.bias)
elif isinstance(m, SAGEConv):
m.reset_parameters()
def encode(self, data):
x_dict = {
'author': self.author_proj(data['author'].x),
'paper': self.paper_proj(data['paper'].x),
}
for conv in self.convs:
x_dict = conv(x_dict, data.edge_index_dict)
return {
'author': self.post_lin(x_dict['author']),
'paper': self.post_lin(x_dict['paper']),
}
def decode(self, z_dict, edge_label_index):
return self.decoder(z_dict['author'], z_dict['paper'], edge_label_index)
# ── Hard negative sampling ───────────────────────────────────────
def sample_hard_negatives(pos_batch_size, num_authors, num_papers,
existing_set, device, pos_src=None):
"""Mixed negative sampling: 50% random, 25% popular, 25% co-author papers."""
neg_list = []
n_random = pos_batch_size // 2
n_popular = pos_batch_size // 4
n_coauthor = pos_batch_size - n_random - n_popular
# Random
while len(neg_list) < n_random:
src = np.random.randint(0, num_authors, size=n_random)
dst = np.random.randint(0, num_papers, size=n_random)
for s, d in zip(src, dst):
if (s, d) not in existing_set:
neg_list.append((s, d))
if len(neg_list) >= n_random:
break
# Popular papers
cnt = 0
while len(neg_list) < n_random + n_popular and cnt < n_popular * 5:
cnt += 1
src = np.random.randint(0, num_authors)
dst = popular_papers[np.random.randint(0, len(popular_papers))]
if (src, dst) not in existing_set:
neg_list.append((src, dst))
# Co-author's papers
cnt = 0
while len(neg_list) < pos_batch_size and cnt < n_coauthor * 10:
cnt += 1
src = np.random.randint(0, num_authors)
pool = coauthor_paper_pool.get(src, [])
if pool:
dst = pool[np.random.randint(0, len(pool))]
if (src, dst) not in existing_set:
neg_list.append((src, dst))
# Fill remaining with random
while len(neg_list) < pos_batch_size:
src = np.random.randint(0, num_authors)
dst = np.random.randint(0, num_papers)
if (src, dst) not in existing_set:
neg_list.append((src, dst))
return torch.tensor(neg_list[:pos_batch_size], dtype=torch.long,
device=device).t().contiguous()
# ── Training setup ────────────────────────────────────────────────
def run_experiment(seed, hidden_dim=128, num_layers=3, lr=0.003,
num_epochs=250, use_hard_neg=True):
set_seed(seed)
model = ImprovedHeteroGNN(
data.metadata(),
author_in_dim=author_deg_dim,
paper_in_dim=paper_feat_dim,
hidden_dim=hidden_dim,
num_layers=num_layers,
out_dim=64,
).to(device)
# Weight decay for non-norm parameters
decay_params = []
no_decay_params = []
for name, param in model.named_parameters():
if 'norm' in name or 'bias' in name:
no_decay_params.append(param)
else:
decay_params.append(param)
optimizer = torch.optim.AdamW([
{'params': decay_params, 'weight_decay': 1e-4},
{'params': no_decay_params, 'weight_decay': 0},
], lr=lr)
scheduler = ReduceLROnPlateau(optimizer, mode='max', factor=0.5,
patience=20, min_lr=1e-6)
pos_edge_index = data['author', 'ref', 'paper'].edge_index
existing_train_set = set(map(tuple, train_refs[['source', 'target']].to_numpy().tolist()))
batch_size = min(32768, pos_edge_index.size(1))
best_val_f1 = 0
best_state = None
patience_counter = 0
for epoch in range(num_epochs):
model.train()
optimizer.zero_grad()
# Sample batch
perm = torch.randperm(pos_edge_index.size(1), device=device)[:batch_size]
pos_batch = pos_edge_index[:, perm]
if use_hard_neg:
neg_batch = sample_hard_negatives(
pos_batch.size(1), num_authors, num_papers,
existing_train_set, device,
pos_src=pos_batch[0].cpu().numpy(),
)
else:
# Random only
neg_list = []
while len(neg_list) < pos_batch.size(1):
s = torch.randint(0, num_authors, (pos_batch.size(1),))
d = torch.randint(0, num_papers, (pos_batch.size(1),))
for si, di in zip(s.tolist(), d.tolist()):
if (si, di) not in existing_train_set:
neg_list.append((si, di))
if len(neg_list) >= pos_batch.size(1):
break
neg_batch = torch.tensor(neg_list, dtype=torch.long,
device=device).t().contiguous()
z_dict = model.encode(data)
pos_score = model.decode(z_dict, pos_batch)
neg_score = model.decode(z_dict, neg_batch)
# BCE loss
pos_labels = torch.ones_like(pos_score)
neg_labels = torch.zeros_like(neg_score)
scores = torch.cat([pos_score, neg_score])
labels = torch.cat([pos_labels, neg_labels])
loss = F.binary_cross_entropy_with_logits(scores, labels)
loss.backward()
torch.nn.utils.clip_grad_norm_(model.parameters(), 1.0)
optimizer.step()
# Validation
if epoch % 5 == 0 or epoch == num_epochs - 1:
val_f1, val_auc, val_thresh = evaluate(model, data, val_set, device)
scheduler.step(val_f1)
if val_f1 > best_val_f1:
best_val_f1 = val_f1
best_state = {k: v.cpu().clone() for k, v in model.state_dict().items()}
patience_counter = 0
else:
patience_counter += 1
if epoch % 20 == 0 or epoch == num_epochs - 1:
print(f'Epoch {epoch:03d} | Loss={loss.item():.4f} | '
f'Val F1={val_f1:.4f} AUC={val_auc:.4f} Thresh={val_thresh:.3f} | '
f'Best F1={best_val_f1:.4f}')
if patience_counter >= 30:
print(f'Early stopping at epoch {epoch}')
break
model.load_state_dict(best_state)
return model, best_val_f1
@torch.no_grad()
def evaluate(model, data, val_df, device):
model.eval()
z_dict = model.encode(data)
val_arr = val_df[['source', 'target']].to_numpy(dtype=np.int64)
val_labels = val_df['label'].to_numpy()
edge_idx = torch.as_tensor(val_arr, device=device).t()
scores = model.decoder(
z_dict['author'], z_dict['paper'],
edge_idx,
).sigmoid().cpu().numpy()
# Best F1 threshold
precision, recall, thresholds = precision_recall_curve(val_labels, scores)
f1s = 2 * precision * recall / (precision + recall + 1e-12)
best_idx = np.argmax(f1s)
best_thresh = thresholds[best_idx] if best_idx < len(thresholds) else 0.5
best_f1 = f1s[best_idx]
auc = roc_auc_score(val_labels, scores)
model.train()
return best_f1, auc, best_thresh
# ── Run experiments ───────────────────────────────────────────────
print("\n" + "=" * 60)
print("Experiment 1: Improved GNN with hard negatives")
print("=" * 60)
model1, f1_1 = run_experiment(seed=0, hidden_dim=128, num_layers=3)
print("\n" + "=" * 60)
print("Experiment 2: Improved GNN (seed=42)")
print("=" * 60)
model2, f1_2 = run_experiment(seed=42, hidden_dim=128, num_layers=3)
print(f"\nModel 1 best val F1: {f1_1:.4f}")
print(f"Model 2 best val F1: {f1_2:.4f}")
# ── Generate final submission (ensemble) ──────────────────────────
print("\n" + "=" * 60)
print("Generating ensemble submission...")
print("=" * 60)
@torch.no_grad()
def predict_all(model, data, test_pairs, overlap_set, device):
model.eval()
z_dict = model.encode(data)
# Batch decode to avoid OOM
batch_size = 131072
all_scores = []
n = len(test_pairs)
for start in range(0, n, batch_size):
end = min(start + batch_size, n)
edge_idx = torch.as_tensor(test_pairs[start:end], device=device).t()
batch_scores = model.decoder(
z_dict['author'], z_dict['paper'],
edge_idx,
).sigmoid().cpu().numpy()
all_scores.append(batch_scores)
return np.concatenate(all_scores)
test_arr = np.array(refs_to_pred, dtype=np.int64)
# Ensemble predictions
scores1 = predict_all(model1, data, test_arr, overlap, device)
scores2 = predict_all(model2, data, test_arr, overlap, device)
ensemble_scores = (scores1 + scores2) / 2.0
# Known positives from training β†’ force to 1
known_pos_mask = np.array([tuple(p) in overlap for p in test_arr])
ensemble_scores[known_pos_mask] = 1.0
# Threshold search on a small validation-like subset to pick threshold
# Since we don't have labels for test, use validation set best threshold
val_scores1 = predict_all(model1, data,
val_set[['source', 'target']].to_numpy(dtype=np.int64),
set(), device)
val_scores2 = predict_all(model2, data,
val_set[['source', 'target']].to_numpy(dtype=np.int64),
set(), device)
val_ens = (val_scores1 + val_scores2) / 2.0
val_labels = val_set['label'].to_numpy()
precision, recall, thresholds = precision_recall_curve(val_labels, val_ens)
f1s = 2 * precision * recall / (precision + recall + 1e-12)
best_idx = np.argmax(f1s)
best_thresh = thresholds[best_idx] if best_idx < len(thresholds) else 0.5
val_f1_ens = f1s[best_idx]
print(f"Ensemble val F1: {val_f1_ens:.4f} @ threshold={best_thresh:.4f}")
# Apply threshold
predictions = (ensemble_scores >= best_thresh).astype(int)
print(f"Predicted positive ratio: {predictions.mean():.4f} "
f"({predictions.sum()} / {len(predictions)})")
print(f"Known positives set to 1: {known_pos_mask.sum()}")
# Save
output_path = "/home/lzc/submission_improved.csv"
data_out = [[idx, str(int(p))] for idx, p in enumerate(predictions)]
df = pd.DataFrame(data_out, columns=['Index', 'Predicted'], dtype=object)
df.to_csv(output_path, index=False)
print(f"\nSubmission saved to: {output_path}")
# Also save individual model submissions
for i, (model, scores_i, name) in enumerate([
(model1, scores1, 'model1'),
(model2, scores2, 'model2'),
]):
s = scores_i.copy()
s[known_pos_mask] = 1.0
preds = (s >= best_thresh).astype(int)
out_path = f"/home/lzc/submission_{name}.csv"
data_out = [[idx, str(int(p))] for idx, p in enumerate(preds)]
pd.DataFrame(data_out, columns=['Index', 'Predicted'], dtype=object).to_csv(
out_path, index=False)
print(f" {name} saved to: {out_path}")