DCCF/model.py at main · HKUDS/DCCF · GitHub

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
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
import torch
import numpy as np
import torch_sparse
import torch.nn as nn
import scipy.sparse as sp
import torch.nn.functional as F

class DCCF(nn.Module):
    def __init__(self, data_config, args):
        super(DCCF, self).__init__()

        self.n_users = data_config['n_users']
        self.n_items = data_config['n_items']

        self.plain_adj = data_config['plain_adj']
        self.all_h_list = data_config['all_h_list']
        self.all_t_list = data_config['all_t_list']
        self.A_in_shape = self.plain_adj.tocoo().shape
        self.A_indices = torch.tensor([self.all_h_list, self.all_t_list], dtype=torch.long).cuda()
        self.D_indices = torch.tensor([list(range(self.n_users + self.n_items)), list(range(self.n_users + self.n_items))], dtype=torch.long).cuda()
        self.all_h_list = torch.LongTensor(self.all_h_list).cuda()
        self.all_t_list = torch.LongTensor(self.all_t_list).cuda()
        self.G_indices, self.G_values = self._cal_sparse_adj()

        self.emb_dim = args.embed_size
        self.n_layers = args.n_layers
        self.n_intents = args.n_intents
        self.temp = args.temp

        self.batch_size = args.batch_size
        self.emb_reg = args.emb_reg
        self.cen_reg = args.cen_reg
        self.ssl_reg = args.ssl_reg


        """
        *********************************************************
        Create Model Parameters
        """
        self.user_embedding = nn.Embedding(self.n_users, self.emb_dim)
        self.item_embedding = nn.Embedding(self.n_items, self.emb_dim)

        _user_intent = torch.empty(self.emb_dim, self.n_intents)
        nn.init.xavier_normal_(_user_intent)
        self.user_intent = torch.nn.Parameter(_user_intent, requires_grad=True)
        _item_intent = torch.empty(self.emb_dim, self.n_intents)
        nn.init.xavier_normal_(_item_intent)
        self.item_intent = torch.nn.Parameter(_item_intent, requires_grad=True)

        """
        *********************************************************
        Initialize Weights
        """
        self._init_weight()

    def _init_weight(self):
        nn.init.xavier_normal_(self.user_embedding.weight)
        nn.init.xavier_normal_(self.item_embedding.weight)

    def _cal_sparse_adj(self):

        A_values = torch.ones(size=(len(self.all_h_list), 1)).view(-1).cuda()

        A_tensor = torch_sparse.SparseTensor(row=self.all_h_list, col=self.all_t_list, value=A_values, sparse_sizes=self.A_in_shape).cuda()
        D_values = A_tensor.sum(dim=1).pow(-0.5)

        G_indices, G_values = torch_sparse.spspmm(self.D_indices, D_values, self.A_indices, A_values, self.A_in_shape[0], self.A_in_shape[1], self.A_in_shape[1])
        G_indices, G_values = torch_sparse.spspmm(G_indices, G_values, self.D_indices, D_values, self.A_in_shape[0], self.A_in_shape[1], self.A_in_shape[1])

        return G_indices, G_values

    def _adaptive_mask(self, head_embeddings, tail_embeddings):

        head_embeddings = torch.nn.functional.normalize(head_embeddings)
        tail_embeddings = torch.nn.functional.normalize(tail_embeddings)
        edge_alpha = (torch.sum(head_embeddings * tail_embeddings, dim=1).view(-1) + 1) / 2

        A_tensor = torch_sparse.SparseTensor(row=self.all_h_list, col=self.all_t_list, value=edge_alpha, sparse_sizes=self.A_in_shape).cuda()
        D_scores_inv = A_tensor.sum(dim=1).pow(-1).nan_to_num(0, 0, 0).view(-1)

        G_indices = torch.stack([self.all_h_list, self.all_t_list], dim=0)
        G_values = D_scores_inv[self.all_h_list] * edge_alpha

        return G_indices, G_values

    def inference(self):
        all_embeddings = [torch.concat([self.user_embedding.weight, self.item_embedding.weight], dim=0)]

        gnn_embeddings = []
        int_embeddings = []
        gaa_embeddings = []
        iaa_embeddings = []

        for i in range(0, self.n_layers):

            # Graph-based Message Passing
            gnn_layer_embeddings = torch_sparse.spmm(self.G_indices, self.G_values, self.A_in_shape[0], self.A_in_shape[1], all_embeddings[i])

            # Intent-aware Information Aggregation
            u_embeddings, i_embeddings = torch.split(all_embeddings[i], [self.n_users, self.n_items], 0)
            u_int_embeddings = torch.softmax(u_embeddings @ self.user_intent, dim=1) @ self.user_intent.T
            i_int_embeddings = torch.softmax(i_embeddings @ self.item_intent, dim=1) @ self.item_intent.T
            int_layer_embeddings = torch.concat([u_int_embeddings, i_int_embeddings], dim=0)

            # Adaptive Augmentation
            gnn_head_embeddings = torch.index_select(gnn_layer_embeddings, 0, self.all_h_list)
            gnn_tail_embeddings = torch.index_select(gnn_layer_embeddings, 0, self.all_t_list)
            int_head_embeddings = torch.index_select(int_layer_embeddings, 0, self.all_h_list)
            int_tail_embeddings = torch.index_select(int_layer_embeddings, 0, self.all_t_list)
            G_graph_indices, G_graph_values = self._adaptive_mask(gnn_head_embeddings, gnn_tail_embeddings)
            G_inten_indices, G_inten_values = self._adaptive_mask(int_head_embeddings, int_tail_embeddings)

            gaa_layer_embeddings = torch_sparse.spmm(G_graph_indices, G_graph_values, self.A_in_shape[0], self.A_in_shape[1], all_embeddings[i])
            iaa_layer_embeddings = torch_sparse.spmm(G_inten_indices, G_inten_values, self.A_in_shape[0], self.A_in_shape[1], all_embeddings[i])

            gnn_embeddings.append(gnn_layer_embeddings)
            int_embeddings.append(int_layer_embeddings)
            gaa_embeddings.append(gaa_layer_embeddings)
            iaa_embeddings.append(iaa_layer_embeddings)

            all_embeddings.append(gnn_layer_embeddings + int_layer_embeddings + gaa_layer_embeddings + iaa_layer_embeddings + all_embeddings[i])

        all_embeddings = torch.stack(all_embeddings, dim=1)
        all_embeddings = torch.sum(all_embeddings, dim=1, keepdim=False)

        self.ua_embedding, self.ia_embedding = torch.split(all_embeddings, [self.n_users, self.n_items], 0)

        return gnn_embeddings, int_embeddings, gaa_embeddings, iaa_embeddings

    def cal_ssl_loss(self, users, items, gnn_emb, int_emb, gaa_emb, iaa_emb):
        users = torch.unique(users)
        items = torch.unique(items)

        cl_loss = 0.0

        def cal_loss(emb1, emb2):
            pos_score = torch.exp(torch.sum(emb1 * emb2, dim=1) / self.temp)
            neg_score = torch.sum(torch.exp(torch.mm(emb1, emb2.T) / self.temp), axis=1)
            loss = torch.sum(-torch.log(pos_score / (neg_score + 1e-8) + 1e-8))
            loss /= pos_score.shape[0]
            return loss

        for i in range(len(gnn_emb)):
            u_gnn_embs, i_gnn_embs = torch.split(gnn_emb[i], [self.n_users, self.n_items], 0)
            u_int_embs, i_int_embs = torch.split(int_emb[i], [self.n_users, self.n_items], 0)
            u_gaa_embs, i_gaa_embs = torch.split(gaa_emb[i], [self.n_users, self.n_items], 0)
            u_iaa_embs, i_iaa_embs = torch.split(iaa_emb[i], [self.n_users, self.n_items], 0)

            u_gnn_embs = F.normalize(u_gnn_embs[users], dim=1)
            u_int_embs = F.normalize(u_int_embs[users], dim=1)
            u_gaa_embs = F.normalize(u_gaa_embs[users], dim=1)
            u_iaa_embs = F.normalize(u_iaa_embs[users], dim=1)

            i_gnn_embs = F.normalize(i_gnn_embs[items], dim=1)
            i_int_embs = F.normalize(i_int_embs[items], dim=1)
            i_gaa_embs = F.normalize(i_gaa_embs[items], dim=1)
            i_iaa_embs = F.normalize(i_iaa_embs[items], dim=1)

            cl_loss += cal_loss(u_gnn_embs, u_int_embs)
            cl_loss += cal_loss(u_gnn_embs, u_gaa_embs)
            cl_loss += cal_loss(u_gnn_embs, u_iaa_embs)

            cl_loss += cal_loss(i_gnn_embs, i_int_embs)
            cl_loss += cal_loss(i_gnn_embs, i_gaa_embs)
            cl_loss += cal_loss(i_gnn_embs, i_iaa_embs)

        return cl_loss

    def forward(self, users, pos_items, neg_items):
        users = torch.LongTensor(users).cuda()
        pos_items = torch.LongTensor(pos_items).cuda()
        neg_items = torch.LongTensor(neg_items).cuda()

        gnn_embeddings, int_embeddings, gaa_embeddings, iaa_embeddings = self.inference()

        # bpr
        u_embeddings = self.ua_embedding[users]
        pos_embeddings = self.ia_embedding[pos_items]
        neg_embeddings = self.ia_embedding[neg_items]
        pos_scores = torch.sum(u_embeddings * pos_embeddings, 1)
        neg_scores = torch.sum(u_embeddings * neg_embeddings, 1)
        mf_loss = torch.mean(F.softplus(neg_scores - pos_scores))

        # embeddings
        u_embeddings_pre = self.user_embedding(users)
        pos_embeddings_pre = self.item_embedding(pos_items)
        neg_embeddings_pre = self.item_embedding(neg_items)
        emb_loss = (u_embeddings_pre.norm(2).pow(2) + pos_embeddings_pre.norm(2).pow(2) + neg_embeddings_pre.norm(2).pow(2))
        emb_loss = self.emb_reg * emb_loss

        # intent prototypes
        cen_loss = (self.user_intent.norm(2).pow(2) + self.item_intent.norm(2).pow(2))
        cen_loss = self.cen_reg * cen_loss

        # self-supervise learning
        cl_loss = self.ssl_reg * self.cal_ssl_loss(users, pos_items, gnn_embeddings, int_embeddings, gaa_embeddings, iaa_embeddings)

        return mf_loss, emb_loss, cen_loss, cl_loss

    def predict(self, users):
        u_embeddings = self.ua_embedding[torch.LongTensor(users).cuda()]
        i_embeddings = self.ia_embedding
        batch_ratings = torch.matmul(u_embeddings, i_embeddings.T)
        return batch_ratings