def __init__(self, config, dataset):
super(MacridVAE, self).__init__(config, dataset)
self.layers = config['encoder_hidden_size']
self.embedding_size = config['embedding_size']
self.drop_out = config['drop_out']
self.kfac = config['kfac']
self.tau = config['tau']
self.nogb = config['nogb']
self.anneal_cap = config['anneal_cap']
self.total_anneal_steps = config['total_anneal_steps']
self.regs = config['reg_weights']
self.std = config['std']
self.update = 0
self.history_item_id, self.history_item_value, _ = dataset.history_item_matrix()
self.history_item_id = self.history_item_id.to(self.device)
self.history_item_value = self.history_item_value.to(self.device)
self.encode_layer_dims = [self.n_items] + self.layers + [self.embedding_size * 2]
self.encoder = self.mlp_layers(self.encode_layer_dims)
self.item_embedding = nn.Embedding(self.n_items, self.embedding_size)
self.k_embedding = nn.Embedding(self.kfac, self.embedding_size)
self.l2_loss = EmbLoss()
# parameters initialization
self.apply(xavier_normal_initialization)
def __init__(self, config, dataset):
super(LightGCN, self).__init__(config, dataset)
# load dataset info
self.interaction_matrix = dataset.inter_matrix(form='coo').astype(
np.float32)
# load parameters info
self.latent_dim = config[
'embedding_size'] # int type:the embedding size of lightGCN
self.n_layers = config[
'n_layers'] # int type:the layer num of lightGCN
self.reg_weight = config[
'reg_weight'] # float32 type: the weight decay for l2 normalizaton
# define layers and loss
self.user_embedding = torch.nn.Embedding(num_embeddings=self.n_users,
embedding_dim=self.latent_dim)
self.item_embedding = torch.nn.Embedding(num_embeddings=self.n_items,
embedding_dim=self.latent_dim)
self.mf_loss = BPRLoss()
self.reg_loss = EmbLoss()
# storage variables for full sort evaluation acceleration
self.restore_user_e = None
self.restore_item_e = None
# generate intermediate data
self.norm_adj_matrix = self.get_norm_adj_mat().to(self.device)
# parameters initialization
self.apply(xavier_uniform_initialization)
def __init__(self, config, dataset):
super(NGCF, self).__init__(config, dataset)
# load dataset info
self.interaction_matrix = dataset.inter_matrix(form='coo').astype(
np.float32)
# load parameters info
self.embedding_size = config['embedding_size']
self.hidden_size_list = config['hidden_size_list']
self.hidden_size_list = [self.embedding_size] + self.hidden_size_list
self.node_dropout = config['node_dropout']
self.message_dropout = config['message_dropout']
self.reg_weight = config['reg_weight']
# define layers and loss
self.user_embedding = nn.Embedding(self.n_users, self.embedding_size)
self.item_embedding = nn.Embedding(self.n_items, self.embedding_size)
self.GNNlayers = torch.nn.ModuleList()
for idx, (input_size, output_size) in enumerate(
zip(self.hidden_size_list[:-1], self.hidden_size_list[1:])):
self.GNNlayers.append(BiGNNLayer(input_size, output_size))
self.mf_loss = BPRLoss()
self.reg_loss = EmbLoss()
# storage variables for full sort evaluation acceleration
self.restore_user_e = None
self.restore_item_e = None
# generate intermediate data
self.norm_adj_matrix = self.get_norm_adj_mat().to(self.device)
self.eye_matrix = self.get_eye_mat().to(self.device)
# parameters initialization
self.apply(xavier_normal_initialization)
def __init__(self, config, dataset):
super(TransRec, self).__init__(config, dataset)
# load parameters info
self.embedding_size = config['embedding_size']
# load dataset info
self.n_users = dataset.user_num
self.user_embedding = nn.Embedding(self.n_users,
self.embedding_size,
padding_idx=0)
self.item_embedding = nn.Embedding(self.n_items,
self.embedding_size,
padding_idx=0)
self.bias = nn.Embedding(self.n_items, 1,
padding_idx=0) # Beta popularity bias
self.T = nn.Parameter(torch.zeros(
self.embedding_size)) # average user representation 'global'
self.bpr_loss = BPRLoss()
self.emb_loss = EmbLoss()
self.reg_loss = RegLoss()
# parameters initialization
self.apply(xavier_normal_initialization)
def __init__(self, config, dataset):
super(SpectralCF, self).__init__(config, dataset)
# load parameters info
self.n_layers = config['n_layers']
self.emb_dim = config['embedding_size']
self.reg_weight = config['reg_weight']
# generate intermediate data
# "A_hat = I + L" is equivalent to "A_hat = U U^T + U \Lambda U^T"
self.interaction_matrix = dataset.inter_matrix(
form='coo').astype(np.float32)
I = self.get_eye_mat(self.n_items + self.n_users)
L = self.get_laplacian_matrix()
A_hat = I + L
self.A_hat = A_hat.to(self.device)
# define layers and loss
self.user_embedding = torch.nn.Embedding(
num_embeddings=self.n_users, embedding_dim=self.emb_dim)
self.item_embedding = torch.nn.Embedding(
num_embeddings=self.n_items, embedding_dim=self.emb_dim)
self.filters = torch.nn.ParameterList(
[torch.nn.Parameter(torch.normal(mean=0.01, std=0.02, size=(self.emb_dim, self.emb_dim)).to(self.device),
requires_grad=True)
for _ in range(self.n_layers)])
self.sigmoid = torch.nn.Sigmoid()
self.mf_loss = BPRLoss()
self.reg_loss = EmbLoss()
self.restore_user_e = None
self.restore_item_e = None
# parameters initialization
self.apply(xavier_uniform_initialization)
def __init__(self, config, dataset):
super(TransRecF, self).__init__(config, dataset)
# load parameters info
self.embedding_size = config['embedding_size']
# load dataset info
self.n_users = dataset.user_num
self.user_embedding = nn.Embedding(self.n_users, self.embedding_size, padding_idx=0)
self.item_embedding = nn.Embedding(self.n_items, self.embedding_size, padding_idx=0)
self.bias = nn.Embedding(self.n_items, 1, padding_idx=0) # Beta popularity bias
self.T = nn.Parameter(torch.zeros(self.embedding_size)) # average user representation 'global'
self.selected_features = config['selected_features']
self.hidden_dropout_prob = config['hidden_dropout_prob']
self.pooling_mode = config['pooling_mode']
self.dropout = nn.Dropout(self.hidden_dropout_prob)
self.layer_norm_eps = config['layer_norm_eps']
self.device = config['device']
self.num_feature_field = len(config['selected_features'])
self.bpr_loss = BPRLoss()
self.emb_loss = EmbLoss()
self.reg_loss = RegLoss()
self.feature_embed_layer = FeatureSeqEmbLayer(
dataset, self.embedding_size, self.selected_features, self.pooling_mode, self.device
)
self.LayerNorm = nn.LayerNorm(self.embedding_size, eps=self.layer_norm_eps)
self.concat_layer = nn.Linear(self.embedding_size * (1 + self.num_feature_field), self.embedding_size)
# parameters initialization
self.apply(xavier_normal_initialization)
def __init__(self, config, dataset):
super(DGCF, self).__init__(config, dataset)
# load dataset info
self.interaction_matrix = dataset.inter_matrix(form='coo').astype(
np.float32)
# load parameters info
self.embedding_size = config['embedding_size']
self.n_factors = config['n_factors']
self.n_iterations = config['n_iterations']
self.n_layers = config['n_layers']
self.reg_weight = config['reg_weight']
self.cor_weight = config['cor_weight']
n_batch = dataset.dataset.inter_num // self.batch_size + 1
self.cor_batch_size = int(
max(self.n_users / n_batch, self.n_items / n_batch))
# ensure embedding can be divided into <n_factors> intent
assert self.embedding_size % self.n_factors == 0
# generate intermediate data
row = self.interaction_matrix.row.tolist()
col = self.interaction_matrix.col.tolist()
col = [item_index + self.n_users for item_index in col]
all_h_list = row + col # row.extend(col)
all_t_list = col + row # col.extend(row)
num_edge = len(all_h_list)
edge_ids = range(num_edge)
self.all_h_list = torch.LongTensor(all_h_list).to(self.device)
self.all_t_list = torch.LongTensor(all_t_list).to(self.device)
self.edge2head = torch.LongTensor([all_h_list,
edge_ids]).to(self.device)
self.head2edge = torch.LongTensor([edge_ids,
all_h_list]).to(self.device)
self.tail2edge = torch.LongTensor([edge_ids,
all_t_list]).to(self.device)
val_one = torch.ones_like(self.all_h_list).float().to(self.device)
num_node = self.n_users + self.n_items
self.edge2head_mat = self._build_sparse_tensor(self.edge2head, val_one,
(num_node, num_edge))
self.head2edge_mat = self._build_sparse_tensor(self.head2edge, val_one,
(num_edge, num_node))
self.tail2edge_mat = self._build_sparse_tensor(self.tail2edge, val_one,
(num_edge, num_node))
self.num_edge = num_edge
self.num_node = num_node
# define layers and loss
self.user_embedding = nn.Embedding(self.n_users, self.embedding_size)
self.item_embedding = nn.Embedding(self.n_items, self.embedding_size)
self.softmax = torch.nn.Softmax(dim=1)
self.mf_loss = BPRLoss()
self.reg_loss = EmbLoss()
self.restore_user_e = None
self.restore_item_e = None
# parameters initialization
self.apply(xavier_normal_initialization)
def __init__(self, config, dataset):
super(RippleNet, self).__init__(config, dataset)
# load dataset info
self.LABEL = config['LABEL_FIELD']
# load parameters info
self.embedding_size = config['embedding_size']
self.kg_weight = config['kg_weight']
self.reg_weight = config['reg_weight']
self.n_hop = config['n_hop']
self.n_memory = config['n_memory']
self.interaction_matrix = dataset.inter_matrix(form='coo').astype(
np.float32)
head_entities = dataset.head_entities.tolist()
tail_entities = dataset.tail_entities.tolist()
relations = dataset.relations.tolist()
kg = {}
for i in range(len(head_entities)):
head_ent = head_entities[i]
tail_ent = tail_entities[i]
relation = relations[i]
kg.setdefault(head_ent, [])
kg[head_ent].append((tail_ent, relation))
self.kg = kg
users = self.interaction_matrix.row.tolist()
items = self.interaction_matrix.col.tolist()
user_dict = {}
for i in range(len(users)):
user = users[i]
item = items[i]
user_dict.setdefault(user, [])
user_dict[user].append(item)
self.user_dict = user_dict
self.ripple_set = self._build_ripple_set()
# define layers and loss
self.entity_embedding = nn.Embedding(self.n_entities,
self.embedding_size)
self.relation_embedding = nn.Embedding(
self.n_relations, self.embedding_size * self.embedding_size)
self.transform_matrix = nn.Linear(self.embedding_size,
self.embedding_size,
bias=False)
self.softmax = torch.nn.Softmax(dim=1)
self.sigmoid = torch.nn.Sigmoid()
self.rec_loss = BPRLoss()
self.l2_loss = EmbLoss()
self.loss = nn.BCEWithLogitsLoss()
# parameters initialization
self.apply(xavier_normal_initialization)
self.other_parameter_name = ['ripple_set']
def __init__(self, config, dataset):
super(KGNNLS, self).__init__(config, dataset)
# load parameters info
self.embedding_size = config['embedding_size']
self.neighbor_sample_size = config['neighbor_sample_size']
self.aggregator_class = config['aggregator'] # which aggregator to use
# number of iterations when computing entity representation
self.n_iter = config['n_iter']
self.reg_weight = config['reg_weight'] # weight of l2 regularization
# weight of label Smoothness regularization
self.ls_weight = config['ls_weight']
# define embedding
self.user_embedding = nn.Embedding(self.n_users, self.embedding_size)
self.entity_embedding = nn.Embedding(self.n_entities,
self.embedding_size)
self.relation_embedding = nn.Embedding(self.n_relations + 1,
self.embedding_size)
# sample neighbors and construct interaction table
kg_graph = dataset.kg_graph(form='coo', value_field='relation_id')
adj_entity, adj_relation = self.construct_adj(kg_graph)
self.adj_entity, self.adj_relation = adj_entity.to(
self.device), adj_relation.to(self.device)
inter_feat = dataset.dataset.inter_feat.values
pos_users = torch.from_numpy(inter_feat[:, 0])
pos_items = torch.from_numpy(inter_feat[:, 1])
pos_label = torch.ones(pos_items.shape)
pos_interaction_table, self.offset = self.get_interaction_table(
pos_users, pos_items, pos_label)
self.interaction_table = self.sample_neg_interaction(
pos_interaction_table, self.offset)
# define function
self.softmax = nn.Softmax(dim=-1)
self.linear_layers = torch.nn.ModuleList()
for i in range(self.n_iter):
self.linear_layers.append(
nn.Linear(
self.embedding_size
if not self.aggregator_class == 'concat' else
self.embedding_size * 2, self.embedding_size))
self.ReLU = nn.ReLU()
self.Tanh = nn.Tanh()
self.bce_loss = nn.BCEWithLogitsLoss()
self.l2_loss = EmbLoss()
# parameters initialization
self.apply(xavier_normal_initialization)
def __init__(self, config, dataset):
super(KGAT, self).__init__(config, dataset)
# load dataset info
self.ckg = dataset.ckg_graph(form='dgl', value_field='relation_id')
self.all_hs = torch.LongTensor(
dataset.ckg_graph(form='coo',
value_field='relation_id').row).to(self.device)
self.all_ts = torch.LongTensor(
dataset.ckg_graph(form='coo',
value_field='relation_id').col).to(self.device)
self.all_rs = torch.LongTensor(
dataset.ckg_graph(form='coo',
value_field='relation_id').data).to(self.device)
self.matrix_size = torch.Size(
[self.n_users + self.n_entities, self.n_users + self.n_entities])
# load parameters info
self.embedding_size = config['embedding_size']
self.kg_embedding_size = config['kg_embedding_size']
self.layers = [self.embedding_size] + config['layers']
self.aggregator_type = config['aggregator_type']
self.mess_dropout = config['mess_dropout']
self.reg_weight = config['reg_weight']
# generate intermediate data
self.A_in = self.init_graph(
) # init the attention matrix by the structure of ckg
# define layers and loss
self.user_embedding = nn.Embedding(self.n_users, self.embedding_size)
self.entity_embedding = nn.Embedding(self.n_entities,
self.embedding_size)
self.relation_embedding = nn.Embedding(self.n_relations,
self.kg_embedding_size)
self.trans_w = nn.Embedding(
self.n_relations, self.embedding_size * self.kg_embedding_size)
self.aggregator_layers = nn.ModuleList()
for idx, (input_dim, output_dim) in enumerate(
zip(self.layers[:-1], self.layers[1:])):
self.aggregator_layers.append(
Aggregator(input_dim, output_dim, self.mess_dropout,
self.aggregator_type))
self.tanh = nn.Tanh()
self.mf_loss = BPRLoss()
self.reg_loss = EmbLoss()
self.restore_user_e = None
self.restore_entity_e = None
# parameters initialization
self.apply(xavier_normal_initialization)
self.other_parameter_name = ['restore_user_e', 'restore_entity_e']
def __init__(self, config, dataset):
super(NCL, self).__init__(config, dataset)
# load dataset info
self.interaction_matrix = dataset.inter_matrix(form='coo').astype(
np.float32)
# load parameters info
self.latent_dim = config[
'embedding_size'] # int type: the embedding size of the base model
self.n_layers = config[
'n_layers'] # int type: the layer num of the base model
self.reg_weight = config[
'reg_weight'] # float32 type: the weight decay for l2 normalization
self.ssl_temp = config['ssl_temp']
self.ssl_reg = config['ssl_reg']
self.hyper_layers = config['hyper_layers']
self.alpha = config['alpha']
self.proto_reg = config['proto_reg']
self.k = config['num_clusters']
# define layers and loss
self.user_embedding = torch.nn.Embedding(num_embeddings=self.n_users,
embedding_dim=self.latent_dim)
self.item_embedding = torch.nn.Embedding(num_embeddings=self.n_items,
embedding_dim=self.latent_dim)
self.mf_loss = BPRLoss()
self.reg_loss = EmbLoss()
# storage variables for full sort evaluation acceleration
self.restore_user_e = None
self.restore_item_e = None
self.norm_adj_mat = self.get_norm_adj_mat().to(self.device)
# parameters initialization
self.apply(xavier_uniform_initialization)
self.other_parameter_name = ['restore_user_e', 'restore_item_e']
self.user_centroids = None
self.user_2cluster = None
self.item_centroids = None
self.item_2cluster = None
def __init__(self, config, dataset):
super(GCSAN, self).__init__(config, dataset)
# load parameters info
self.n_layers = config['n_layers']
self.n_heads = config['n_heads']
self.hidden_size = config['hidden_size'] # same as embedding_size
self.inner_size = config[
'inner_size'] # the dimensionality in feed-forward layer
self.hidden_dropout_prob = config['hidden_dropout_prob']
self.attn_dropout_prob = config['attn_dropout_prob']
self.hidden_act = config['hidden_act']
self.layer_norm_eps = config['layer_norm_eps']
self.step = config['step']
self.device = config['device']
self.weight = config['weight']
self.reg_weight = config['reg_weight']
self.loss_type = config['loss_type']
self.initializer_range = config['initializer_range']
# define layers and loss
self.item_embedding = nn.Embedding(self.n_items,
self.hidden_size,
padding_idx=0)
self.gnn = GNN(self.hidden_size, self.step)
self.self_attention = TransformerEncoder(
n_layers=self.n_layers,
n_heads=self.n_heads,
hidden_size=self.hidden_size,
inner_size=self.inner_size,
hidden_dropout_prob=self.hidden_dropout_prob,
attn_dropout_prob=self.attn_dropout_prob,
hidden_act=self.hidden_act,
layer_norm_eps=self.layer_norm_eps)
self.reg_loss = EmbLoss()
if self.loss_type == 'BPR':
self.loss_fct = BPRLoss()
elif self.loss_type == 'CE':
self.loss_fct = nn.CrossEntropyLoss()
else:
raise NotImplementedError(
"Make sure 'loss_type' in ['BPR', 'CE']!")
# parameters initialization
self.apply(self._init_weights)
def __init__(self, config, dataset):
super(KGCN, self).__init__(config, dataset)
# load parameters info
self.embedding_size = config['embedding_size']
# number of iterations when computing entity representation
self.n_iter = config['n_iter']
self.aggregator_class = config['aggregator'] # which aggregator to use
self.reg_weight = config['reg_weight'] # weight of l2 regularization
self.neighbor_sample_size = config['neighbor_sample_size']
# define embedding
self.user_embedding = nn.Embedding(self.n_users, self.embedding_size)
self.entity_embedding = nn.Embedding(self.n_entities,
self.embedding_size)
self.relation_embedding = nn.Embedding(self.n_relations + 1,
self.embedding_size)
# sample neighbors
kg_graph = dataset.kg_graph(form='coo', value_field='relation_id')
adj_entity, adj_relation = self.construct_adj(kg_graph)
self.adj_entity, self.adj_relation = adj_entity.to(
self.device), adj_relation.to(self.device)
# define function
self.softmax = nn.Softmax(dim=-1)
self.linear_layers = torch.nn.ModuleList()
for i in range(self.n_iter):
self.linear_layers.append(
nn.Linear(
self.embedding_size
if not self.aggregator_class == 'concat' else
self.embedding_size * 2, self.embedding_size))
self.ReLU = nn.ReLU()
self.Tanh = nn.Tanh()
self.bce_loss = nn.BCEWithLogitsLoss()
self.l2_loss = EmbLoss()
# parameters initialization
self.apply(xavier_normal_initialization)
self.other_parameter_name = ['adj_entity', 'adj_relation']
def __init__(self, config, dataset):
super(SGL, self).__init__(config, dataset)
self._user = dataset.inter_feat[dataset.uid_field]
self._item = dataset.inter_feat[dataset.iid_field]
self.embed_dim = config["embedding_size"]
self.n_layers = int(config["n_layers"])
self.type = config["type"]
self.drop_ratio = config["drop_ratio"]
self.ssl_tau = config["ssl_tau"]
self.reg_weight = config["reg_weight"]
self.ssl_weight = config["ssl_weight"]
self.user_embedding = torch.nn.Embedding(self.n_users, self.embed_dim)
self.item_embedding = torch.nn.Embedding(self.n_items, self.embed_dim)
self.reg_loss = EmbLoss()
self.train_graph = self.csr2tensor(
self.create_adjust_matrix(is_sub=False))
self.restore_user_e = None
self.restore_item_e = None
self.apply(xavier_uniform_initialization)
self.other_parameter_name = ['restore_user_e', 'restore_item_e']
def __init__(self, config, dataset):
super(CKE, self).__init__(config, dataset)
# load parameters info
self.embedding_size = config['embedding_size']
self.kg_embedding_size = config['kg_embedding_size']
self.reg_weights = config['reg_weights']
# define layers and loss
self.user_embedding = nn.Embedding(self.n_users, self.embedding_size)
self.item_embedding = nn.Embedding(self.n_items, self.embedding_size)
self.entity_embedding = nn.Embedding(self.n_entities, self.embedding_size)
self.relation_embedding = nn.Embedding(self.n_relations, self.kg_embedding_size)
self.trans_w = nn.Embedding(self.n_relations, self.embedding_size * self.kg_embedding_size)
self.rec_loss = BPRLoss()
self.kg_loss = BPRLoss()
self.reg_loss = EmbLoss()
# parameters initialization
self.apply(xavier_normal_initialization)
