# -*- coding: utf-8 -*-
# @Time : 2020/12/8
# @Author : Yihong Guo
# @Email : gyihong@hotmail.com
r"""
LINE
################################################
Reference:
Jian Tang et al. "LINE: Large-scale Information Network Embedding." in WWW 2015.
Reference code:
https://github.com/shenweichen/GraphEmbedding
"""
import random
import numpy as np
import torch
import torch.nn as nn
from recbole.model.abstract_recommender import GeneralRecommender
from recbole.model.init import xavier_normal_initialization
from recbole.utils import InputType
[docs]class NegSamplingLoss(nn.Module):
def __init__(self):
super(NegSamplingLoss, self).__init__()
[docs] def forward(self, score, sign):
return -torch.mean(torch.sigmoid(sign * score))
[docs]class LINE(GeneralRecommender):
r"""LINE is a graph embedding model.
We implement the model to train users and items embedding for recommendation.
"""
input_type = InputType.PAIRWISE
def __init__(self, config, dataset):
super(LINE, self).__init__(config, dataset)
self.embedding_size = config['embedding_size']
self.order = config['order']
self.second_order_loss_weight = config['second_order_loss_weight']
self.interaction_feat = dataset.inter_feat
self.user_embedding = nn.Embedding(self.n_users, self.embedding_size)
self.item_embedding = nn.Embedding(self.n_items, self.embedding_size)
if self.order == 2:
self.user_context_embedding = nn.Embedding(self.n_users, self.embedding_size)
self.item_context_embedding = nn.Embedding(self.n_items, self.embedding_size)
self.loss_fct = NegSamplingLoss()
self.used_ids = self.get_used_ids()
self.random_list = self.get_user_id_list()
np.random.shuffle(self.random_list)
self.random_pr = 0
self.random_list_length = len(self.random_list)
self.apply(xavier_normal_initialization)
[docs] def get_used_ids(self):
cur = np.array([set() for _ in range(self.n_items)])
for uid, iid in zip(self.interaction_feat[self.USER_ID].numpy(), self.interaction_feat[self.ITEM_ID].numpy()):
cur[iid].add(uid)
return cur
[docs] def sampler(self, key_ids):
key_ids = np.array(key_ids.cpu())
key_num = len(key_ids)
total_num = key_num
value_ids = np.zeros(total_num, dtype=np.int64)
check_list = np.arange(total_num)
key_ids = np.tile(key_ids, 1)
while len(check_list) > 0:
value_ids[check_list] = self.random_num(len(check_list))
check_list = np.array([
i for i, used, v in zip(check_list, self.used_ids[key_ids[check_list]], value_ids[check_list])
if v in used
])
return torch.tensor(value_ids, device=self.device)
[docs] def random_num(self, num):
value_id = []
self.random_pr %= self.random_list_length
while True:
if self.random_pr + num <= self.random_list_length:
value_id.append(self.random_list[self.random_pr:self.random_pr + num])
self.random_pr += num
break
else:
value_id.append(self.random_list[self.random_pr:])
num -= self.random_list_length - self.random_pr
self.random_pr = 0
np.random.shuffle(self.random_list)
return np.concatenate(value_id)
[docs] def get_user_id_list(self):
return np.arange(1, self.n_users)
[docs] def forward(self, h, t):
h_embedding = self.user_embedding(h)
t_embedding = self.item_embedding(t)
return torch.sum(h_embedding.mul(t_embedding), dim=1)
[docs] def context_forward(self, h, t, field):
if field == "uu":
h_embedding = self.user_embedding(h)
t_embedding = self.item_context_embedding(t)
else:
h_embedding = self.item_embedding(h)
t_embedding = self.user_context_embedding(t)
return torch.sum(h_embedding.mul(t_embedding), dim=1)
[docs] def calculate_loss(self, interaction):
user = interaction[self.USER_ID]
pos_item = interaction[self.ITEM_ID]
neg_item = interaction[self.NEG_ITEM_ID]
score_pos = self.forward(user, pos_item)
ones = torch.ones(len(score_pos), device=self.device)
if self.order == 1:
if random.random() < 0.5:
score_neg = self.forward(user, neg_item)
else:
neg_user = self.sampler(pos_item)
score_neg = self.forward(neg_user, pos_item)
return self.loss_fct(ones, score_pos) + self.loss_fct(-1 * ones, score_neg)
else:
# randomly train i-i relation and u-u relation with u-i relation
if random.random() < 0.5:
score_neg = self.forward(user, neg_item)
score_pos_con = self.context_forward(user, pos_item, 'uu')
score_neg_con = self.context_forward(user, neg_item, 'uu')
else:
# sample negative user for item
neg_user = self.sampler(pos_item)
score_neg = self.forward(neg_user, pos_item)
score_pos_con = self.context_forward(pos_item, user, 'ii')
score_neg_con = self.context_forward(pos_item, neg_user, 'ii')
return self.loss_fct(ones, score_pos) \
+ self.loss_fct(-1 * ones, score_neg) \
+ self.loss_fct(ones, score_pos_con) * self.second_order_loss_weight \
+ self.loss_fct(-1 * ones, score_neg_con) * self.second_order_loss_weight
[docs] def predict(self, interaction):
user = interaction[self.USER_ID]
item = interaction[self.ITEM_ID]
scores = self.forward(user, item)
return scores
[docs] def full_sort_predict(self, interaction):
user = interaction[self.USER_ID]
# get user embedding from storage variable
u_embeddings = self.user_embedding(user)
i_embedding = self.item_embedding.weight
# dot with all item embedding to accelerate
scores = torch.matmul(u_embeddings, i_embedding.transpose(0, 1))
return scores.view(-1)