推荐系统（七）—— 图神经网络与社交推荐

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31

import torch
import torch.nn as nn
import torch.nn.functional as F
from torch_geometric.utils import add_self_loops, degree


class GCNLayer(nn.Module):
    """单层图卷积。"""

    def __init__(self, in_channels, out_channels):
        super().__init__()
        self.linear = nn.Linear(in_channels, out_channels)

    def forward(self, x, edge_index):
        # 1. 加自环，让每个节点也听自己
        edge_index, _ = add_self_loops(edge_index, num_nodes=x.size(0))

        # 2. 对称归一化系数
        row, col = edge_index
        deg = degree(col, x.size(0), dtype=x.dtype)
        deg_inv_sqrt = deg.pow(-0.5)
        deg_inv_sqrt[deg_inv_sqrt == float('inf')] = 0
        norm = deg_inv_sqrt[row] * deg_inv_sqrt[col]

        # 3. 线性变换
        x = self.linear(x)

        # 4. scatter-add 把归一化后的邻居特征聚合到目标节点
        out = torch.zeros_like(x)
        out.scatter_add_(0, col.unsqueeze(1).expand_as(x), norm.unsqueeze(1) * x[row])
        return out

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33

class GCN(nn.Module):
    """多层 GCN。"""

    def __init__(self, num_nodes, in_channels, hidden_channels,
                 out_channels, num_layers=2, dropout=0.5):
        super().__init__()
        self.num_layers = num_layers
        self.dropout = dropout

        # 节点没有原始特征时，用 ID 嵌入兜底
        self.embedding = None
        if in_channels == 0:
            self.embedding = nn.Embedding(num_nodes, hidden_channels)
            in_channels = hidden_channels

        self.convs = nn.ModuleList()
        if num_layers == 1:
            self.convs.append(GCNLayer(in_channels, out_channels))
        else:
            self.convs.append(GCNLayer(in_channels, hidden_channels))
            for _ in range(num_layers - 2):
                self.convs.append(GCNLayer(hidden_channels, hidden_channels))
            self.convs.append(GCNLayer(hidden_channels, out_channels))

    def forward(self, x, edge_index):
        if self.embedding is not None:
            x = self.embedding.weight
        for i, conv in enumerate(self.convs):
            x = conv(x, edge_index)
            if i < self.num_layers - 1:
                x = F.relu(x)
                x = F.dropout(x, p=self.dropout, training=self.training)
        return x

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38

import torch
import torch.nn as nn
import torch.nn.functional as F
from torch_geometric.nn import MessagePassing
from torch_geometric.utils import add_self_loops, softmax


class GATLayer(MessagePassing):
    """带多头注意力的 GAT 层。"""

    def __init__(self, in_channels, out_channels, heads=1,
                 dropout=0.0, concat=True):
        super().__init__(aggr='add', node_dim=0)
        self.in_channels = in_channels
        self.out_channels = out_channels
        self.heads = heads
        self.concat = concat
        self.dropout = dropout

        self.lin = nn.Linear(in_channels, heads * out_channels, bias=False)
        self.att = nn.Parameter(torch.empty(1, heads, 2 * out_channels))
        nn.init.xavier_uniform_(self.lin.weight)
        nn.init.xavier_uniform_(self.att)

    def forward(self, x, edge_index):
        edge_index, _ = add_self_loops(edge_index, num_nodes=x.size(0))
        # 投到多头空间： [N, heads, out_channels]
        x = self.lin(x).view(-1, self.heads, self.out_channels)
        out = self.propagate(edge_index, x=x)
        return out.view(-1, self.heads * self.out_channels) if self.concat else out.mean(dim=1)

    def message(self, x_i, x_j, index):
        # x_i 是目标节点， x_j 是源节点
        alpha = (torch.cat([x_i, x_j], dim=-1) * self.att).sum(dim=-1)
        alpha = F.leaky_relu(alpha, negative_slope=0.2)
        alpha = softmax(alpha, index)  # 按目标节点归一化
        alpha = F.dropout(alpha, p=self.dropout, training=self.training)
        return x_j * alpha.unsqueeze(-1)

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24

import torch
import torch.nn as nn
import torch.nn.functional as F
from torch_geometric.nn import MessagePassing
from torch_geometric.utils import add_self_loops


class SAGEConv(MessagePassing):
    """采用均值聚合的 GraphSAGE 层。"""

    def __init__(self, in_channels, out_channels, normalize=False):
        super().__init__(aggr='mean')
        self.normalize = normalize
        self.lin_self = nn.Linear(in_channels, out_channels)   # 自身特征
        self.lin_neigh = nn.Linear(in_channels, out_channels)  # 邻居特征

    def forward(self, x, edge_index):
        edge_index, _ = add_self_loops(edge_index, num_nodes=x.size(0))
        neigh_agg = self.propagate(edge_index, x=x)
        out = self.lin_self(x) + self.lin_neigh(neigh_agg)
        return F.normalize(out, p=2, dim=1) if self.normalize else out

    def message(self, x_j):
        return x_j

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60

import torch
import torch.nn as nn
import numpy as np
from collections import Counter, defaultdict


class PinSageConv(nn.Module):
    """带重要性加权聚合的 PinSage 卷积。"""

    def __init__(self, in_channels, out_channels):
        super().__init__()
        self.aggregator = nn.Sequential(
            nn.Linear(in_channels, out_channels),
            nn.ReLU(),
            nn.Linear(out_channels, out_channels),
        )
        self.combine = nn.Linear(in_channels + out_channels, out_channels)

    def forward(self, x, edge_index, importance_weights=None):
        row, col = edge_index
        neighbor_feats = x[col]
        if importance_weights is not None:
            neighbor_feats = neighbor_feats * importance_weights.unsqueeze(1)

        aggregated = torch.zeros(x.size(0), x.size(1), device=x.device)
        aggregated.scatter_add_(
            0, row.unsqueeze(1).expand_as(neighbor_feats), neighbor_feats
        )
        aggregated = self.aggregator(aggregated)
        return self.combine(torch.cat([x, aggregated], dim=1))


def compute_random_walk_importance(adj_list, num_nodes, num_walks=10,
                                   walk_length=5, top_k=10):
    """用短随机游走估计邻居重要性。"""
    visit_counts = Counter()
    for start in range(num_nodes):
        if start not in adj_list:
            continue
        for _ in range(num_walks):
            current = start
            for _ in range(walk_length):
                neighbors = adj_list.get(current, [])
                if not neighbors:
                    break
                current = np.random.choice(neighbors)
                visit_counts[(start, current)] += 1

    per_source = defaultdict(list)
    for (src, tgt), cnt in visit_counts.items():
        per_source[src].append((tgt, cnt))

    scores = {}
    for src, neighbors in per_source.items():
        neighbors.sort(key=lambda t: t[1], reverse=True)
        top = neighbors[:top_k]
        total = sum(c for _, c in top)
        for tgt, cnt in top:
            scores[(src, tgt)] = cnt / total if total > 0 else 0
    return scores

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53

import torch
import torch.nn as nn
from torch_geometric.nn import MessagePassing
from torch_geometric.utils import degree


class LightGCNLayer(MessagePassing):
    """LightGCN 的一层：归一化邻居聚合，仅此而已。"""

    def __init__(self):
        super().__init__(aggr='add')

    def forward(self, x, edge_index):
        row, col = edge_index
        deg = degree(col, x.size(0), dtype=x.dtype)
        deg_inv_sqrt = deg.pow(-0.5)
        deg_inv_sqrt[deg_inv_sqrt == float('inf')] = 0
        norm = deg_inv_sqrt[row] * deg_inv_sqrt[col]
        return self.propagate(edge_index, x=x, norm=norm)

    def message(self, x_j, norm):
        return norm.unsqueeze(1) * x_j


class LightGCN(nn.Module):
    """完整的 LightGCN 协同过滤模型。"""

    def __init__(self, num_users, num_items, embedding_dim=64, num_layers=3):
        super().__init__()
        self.num_users = num_users
        self.num_items = num_items
        self.num_layers = num_layers

        self.user_embedding = nn.Embedding(num_users, embedding_dim)
        self.item_embedding = nn.Embedding(num_items, embedding_dim)
        nn.init.normal_(self.user_embedding.weight, std=0.1)
        nn.init.normal_(self.item_embedding.weight, std=0.1)

        self.convs = nn.ModuleList([LightGCNLayer() for _ in range(num_layers)])

    def forward(self, edge_index):
        x = torch.cat([self.user_embedding.weight, self.item_embedding.weight], dim=0)
        layer_embeddings = [x]  # 第 0 层 = 原始 embedding
        for conv in self.convs:
            x = conv(x, edge_index)
            layer_embeddings.append(x)

        # 各层等权组合
        final = torch.stack(layer_embeddings, dim=0).mean(dim=0)
        return final[:self.num_users], final[self.num_users:]

    def predict(self, user_emb, item_emb, user_ids, item_ids):
        return (user_emb[user_ids] * item_emb[item_ids]).sum(dim=1)

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58

import torch
import torch.nn as nn
import torch.nn.functional as F
from torch_geometric.nn import MessagePassing
from torch_geometric.utils import degree


class NGCFLayer(MessagePassing):
    """带特征变换与交互的 NGCF 层。"""

    def __init__(self, in_channels, out_channels, dropout=0.0):
        super().__init__(aggr='add')
        self.W1 = nn.Linear(in_channels, out_channels, bias=False)
        self.W2 = nn.Linear(in_channels, out_channels, bias=False)
        self.W_self = nn.Linear(in_channels, out_channels, bias=False)
        self.dropout = dropout

    def forward(self, x, edge_index):
        row, col = edge_index
        deg = degree(col, x.size(0), dtype=x.dtype)
        deg_inv_sqrt = deg.pow(-0.5)
        deg_inv_sqrt[deg_inv_sqrt == float('inf')] = 0
        norm = deg_inv_sqrt[row] * deg_inv_sqrt[col]

        self_emb = self.W_self(x)
        neighbor_emb = self.propagate(edge_index, x=x, norm=norm)
        out = F.leaky_relu(self_emb + neighbor_emb, negative_slope=0.2)
        return F.dropout(out, p=self.dropout, training=self.training)

    def message(self, x_i, x_j, norm):
        # W1 * 邻居 + W2 * (用户 . 物品) —— 特征交互
        msg = self.W1(x_j) + self.W2(x_i * x_j)
        return norm.unsqueeze(1) * msg


class NGCF(nn.Module):
    """神经图协同过滤。"""

    def __init__(self, num_users, num_items, embedding_dim=64,
                 num_layers=3, dropout=0.1):
        super().__init__()
        self.num_users = num_users
        self.num_items = num_items
        self.user_embedding = nn.Embedding(num_users, embedding_dim)
        self.item_embedding = nn.Embedding(num_items, embedding_dim)
        nn.init.normal_(self.user_embedding.weight, std=0.1)
        nn.init.normal_(self.item_embedding.weight, std=0.1)

        self.convs = nn.ModuleList([
            NGCFLayer(embedding_dim, embedding_dim, dropout)
            for _ in range(num_layers)
        ])

    def forward(self, edge_index):
        x = torch.cat([self.user_embedding.weight, self.item_embedding.weight], dim=0)
        for conv in self.convs:
            x = conv(x, edge_index)
        return x[:self.num_users], x[self.num_users:]

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40

import torch
import torch.nn as nn
import torch.nn.functional as F
from torch_geometric.nn import MessagePassing


class SocialGCNLayer(MessagePassing):
    """同时处理用户-物品边与社交边的 GCN 层。"""

    def __init__(self, in_channels, out_channels):
        super().__init__(aggr='mean')
        self.linear = nn.Linear(in_channels, out_channels)

    def forward(self, x, edge_index_ui, edge_index_social, num_users):
        x_ui = self.propagate(edge_index_ui, x=x)
        x_social = self.propagate(edge_index_social, x=x[:num_users])
        x_combined = x_ui.clone()
        x_combined[:num_users] = x_combined[:num_users] + x_social
        return self.linear(x_combined)


class SocialGCN(nn.Module):
    """基于 GCN 的社交推荐。"""

    def __init__(self, num_users, num_items, embedding_dim=64, num_layers=2):
        super().__init__()
        self.num_users = num_users
        self.user_embedding = nn.Embedding(num_users, embedding_dim)
        self.item_embedding = nn.Embedding(num_items, embedding_dim)
        self.convs = nn.ModuleList([
            SocialGCNLayer(embedding_dim, embedding_dim)
            for _ in range(num_layers)
        ])

    def forward(self, edge_index_ui, edge_index_social):
        x = torch.cat([self.user_embedding.weight, self.item_embedding.weight], dim=0)
        for conv in self.convs:
            x = conv(x, edge_index_ui, edge_index_social, self.num_users)
            x = F.relu(x)
        return x[:self.num_users], x[self.num_users:]

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44

import numpy as np
import torch


class NeighborSampler:
    """GraphSAGE 风格的邻居采样器，用于小批量训练。"""

    def __init__(self, edge_index, num_nodes, num_neighbors=(10, 5)):
        """num_neighbors: (一跳采样数, 二跳采样数, …)"""
        self.num_nodes = num_nodes
        self.num_neighbors = num_neighbors

        self.adj = {i: [] for i in range(num_nodes)}
        row, col = edge_index.cpu().numpy()
        for r, c in zip(row, col):
            self.adj[r].append(c)

    def sample(self, target_nodes):
        """为目标节点采样多跳邻域。"""
        current_nodes = set(target_nodes.tolist())
        all_layers = []

        for num_nbrs in self.num_neighbors:
            edges, next_nodes = [], set()
            for node in current_nodes:
                neighbors = self.adj[node]
                if not neighbors:
                    continue
                if len(neighbors) > num_nbrs:
                    sampled = np.random.choice(
                        neighbors, num_nbrs, replace=False
                    ).tolist()
                else:
                    sampled = neighbors
                for nbr in sampled:
                    edges.append([nbr, node])
                    next_nodes.add(nbr)

            if edges:
                all_layers.append(torch.tensor(edges).t().contiguous())
            else:
                all_layers.append(torch.empty((2, 0), dtype=torch.long))
            current_nodes = next_nodes
        return all_layers

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51

import torch
import torch.nn as nn
import torch.optim as optim


class BPRLoss(nn.Module):
    """带 L2 正则的 BPR 损失。"""

    def __init__(self, reg_lambda=1e-4):
        super().__init__()
        self.reg_lambda = reg_lambda

    def forward(self, pos_scores, neg_scores, *embeddings):
        bpr = -torch.log(torch.sigmoid(pos_scores - neg_scores) + 1e-10).mean()
        reg = self.reg_lambda * sum(emb.norm(2).pow(2) for emb in embeddings)
        return bpr + reg


def train_lightgcn(model, edge_index, train_pairs, num_items,
                   num_epochs=100, lr=0.001, batch_size=1024):
    """训练 LightGCN 。 train_pairs: [user, pos_item] 张量。"""
    optimizer = optim.Adam(model.parameters(), lr=lr)
    criterion = BPRLoss()

    for epoch in range(num_epochs):
        model.train()
        perm = torch.randperm(len(train_pairs))
        total_loss, num_batches = 0.0, 0

        for start in range(0, len(train_pairs), batch_size):
            batch = train_pairs[perm[start:start + batch_size]]
            user_ids, pos_items = batch[:, 0], batch[:, 1]
            neg_items = torch.randint(0, num_items, (len(batch),))

            optimizer.zero_grad()
            user_emb, item_emb = model(edge_index)

            pos_scores = (user_emb[user_ids] * item_emb[pos_items]).sum(1)
            neg_scores = (user_emb[user_ids] * item_emb[neg_items]).sum(1)

            loss = criterion(pos_scores, neg_scores,
                             user_emb[user_ids],
                             item_emb[pos_items],
                             item_emb[neg_items])
            loss.backward()
            optimizer.step()
            total_loss += loss.item()
            num_batches += 1

        if (epoch + 1) % 10 == 0:
            print(f"Epoch {epoch+1}/{num_epochs}, Loss: {total_loss/num_batches:.4f}")

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32

import numpy as np
import torch


def evaluate_topk(model, edge_index, test_user_items, train_user_items, k=10):
    """计算 Recall@K 和 NDCG@K 。"""
    model.eval()
    recalls, ndcgs = [], []

    with torch.no_grad():
        user_emb, item_emb = model(edge_index)
        for user_id, test_items in test_user_items.items():
            scores = (user_emb[user_id] * item_emb).sum(dim=1)

            # 屏蔽训练集物品
            train_items = train_user_items.get(user_id, [])
            scores[train_items] = -float('inf')

            _, top_k = torch.topk(scores, k)
            top_k = set(top_k.cpu().numpy())

            hits = len(top_k & set(test_items))
            recalls.append(hits / len(test_items) if test_items else 0)

            # NDCG：奖励排位靠前的命中
            dcg = sum(1.0 / np.log2(idx + 2)
                      for idx, item in enumerate(top_k) if item in test_items)
            idcg = sum(1.0 / np.log2(idx + 2)
                       for idx in range(min(k, len(test_items))))
            ndcgs.append(dcg / idcg if idcg > 0 else 0)

    return np.mean(recalls), np.mean(ndcgs)

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49

import torch
import numpy as np
from sklearn.model_selection import train_test_split

# --- 1. 合成类 MovieLens 数据 ---
num_users, num_items = 1000, 2000
num_interactions = 10000

users = np.random.randint(0, num_users, num_interactions)
items = np.random.randint(0, num_items, num_interactions)
train_u, test_u, train_i, test_i = train_test_split(
    users, items, test_size=0.2, random_state=42
)

# 二部图边索引（无向）
edge_src = np.concatenate([train_u, train_i + num_users])
edge_dst = np.concatenate([train_i + num_users, train_u])
edge_index = torch.tensor([edge_src, edge_dst], dtype=torch.long)

# --- 2. 训练 LightGCN ---
model = LightGCN(num_users, num_items, embedding_dim=64, num_layers=3)
optimizer = torch.optim.Adam(model.parameters(), lr=0.001)
criterion = BPRLoss()

train_pairs = torch.tensor(np.stack([train_u, train_i], axis=1), dtype=torch.long)

for epoch in range(50):
    model.train()
    perm = torch.randperm(len(train_pairs))
    for start in range(0, len(train_pairs), 512):
        batch = train_pairs[perm[start:start + 512]]
        user_ids, pos_items = batch[:, 0], batch[:, 1]
        neg_items = torch.randint(0, num_items, (len(batch),))

        optimizer.zero_grad()
        user_emb, item_emb = model(edge_index)
        pos_scores = (user_emb[user_ids] * item_emb[pos_items]).sum(1)
        neg_scores = (user_emb[user_ids] * item_emb[neg_items]).sum(1)
        loss = criterion(pos_scores, neg_scores,
                         user_emb[user_ids],
                         item_emb[pos_items],
                         item_emb[neg_items])
        loss.backward()
        optimizer.step()

    if (epoch + 1) % 10 == 0:
        print(f"Epoch {epoch+1}, Loss: {loss.item():.4f}")

print("训练完成！")

1
2
3
4
5
6

def temporal_aggregate(neighbor_embs, timestamps, current_time, decay_rate=0.1):
    """按时近度给邻居加权。"""
    time_diffs = current_time - timestamps
    weights = torch.exp(-decay_rate * time_diffs)
    weights = weights / weights.sum()
    return (neighbor_embs * weights.unsqueeze(1)).sum(dim=0)