# -*- coding: utf-8 -*-
# @Time : 2020/10/2
# @Author : Jingsen Zhang
# @Email : zhangjingsen@ruc.edu.cn
r"""
NextItNet
################################################
Reference:
Fajie Yuan et al., "A Simple Convolutional Generative Network for Next Item Recommendation" in WSDM 2019.
Reference code:
- https://github.com/fajieyuan/nextitnet
- https://github.com/initlisk/nextitnet_pytorch
"""
import numpy as np
import torch
from torch import nn
from torch.nn import functional as F
from torch.nn.init import uniform_, xavier_normal_, constant_
from recbole.model.loss import RegLoss, BPRLoss
from recbole.model.abstract_recommender import SequentialRecommender
[docs]class NextItNet(SequentialRecommender):
r"""The network architecture of the NextItNet model is formed of a stack of holed convolutional layers, which can
efficiently increase the receptive fields without relying on the pooling operation.
Also residual block structure is used to ease the optimization for much deeper networks.
Note:
As paper said, for comparison purpose, we only predict the next one item in our evaluation,
and then stop the generating process. Although the number of parameters in residual block (a) is less
than it in residual block (b), the performance of b is better than a.
So in our model, we use residual block (b).
"""
def __init__(self, config, dataset):
super(NextItNet, self).__init__(config, dataset)
# load parameters info
self.embedding_size = config['embedding_size']
self.residual_channels = config['embedding_size']
self.block_num = config['block_num']
self.dilations = config['dilations'] * self.block_num
self.kernel_size = config['kernel_size']
self.reg_weight = config['reg_weight']
self.loss_type = config['loss_type']
# define layers and loss
self.item_embedding = nn.Embedding(self.n_items, self.embedding_size, padding_idx=0)
# residual blocks dilations in blocks:[1,2,4,8,1,2,4,8,...]
rb = [ResidualBlock_b(self.residual_channels, self.residual_channels, kernel_size=self.kernel_size,
dilation=dilation) for dilation in self.dilations]
self.residual_blocks = nn.Sequential(*rb)
# fully-connected layer
self.final_layer = nn.Linear(self.residual_channels, self.embedding_size)
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']!")
self.reg_loss = RegLoss()
# parameters initialization
self.apply(self._init_weights)
def _init_weights(self, module):
if isinstance(module, nn.Embedding):
stdv = np.sqrt(1. / self.n_items)
uniform_(module.weight.data, -stdv, stdv)
elif isinstance(module, nn.Linear):
xavier_normal_(module.weight.data)
if module.bias is not None:
constant_(module.bias.data, 0.1)
[docs] def forward(self, item_seq):
item_seq_emb = self.item_embedding(item_seq) # [batch_size, seq_len, embed_size]
# Residual locks
dilate_outputs = self.residual_blocks(item_seq_emb)
hidden = dilate_outputs[:, -1, :].view(-1, self.residual_channels) # [batch_size, embed_size]
seq_output = self.final_layer(hidden) # [batch_size, embedding_size]
return seq_output
[docs] def reg_loss_rb(self):
r"""
L2 loss on residual blocks
"""
loss_rb = 0
if self.reg_weight > 0.0:
for name, parm in self.residual_blocks.named_parameters():
if name.endswith('weight'):
loss_rb += torch.norm(parm,2)
return self.reg_weight * loss_rb
[docs] def calculate_loss(self, interaction):
item_seq = interaction[self.ITEM_SEQ]
# item_seq_len = interaction[self.ITEM_SEQ_LEN]
seq_output = self.forward(item_seq)
pos_items = interaction[self.POS_ITEM_ID]
if self.loss_type == 'BPR':
neg_items = interaction[self.NEG_ITEM_ID]
pos_items_emb = self.item_embedding(pos_items)
neg_items_emb = self.item_embedding(neg_items)
pos_score = torch.sum(seq_output * pos_items_emb, dim=-1) # [B]
neg_score = torch.sum(seq_output * neg_items_emb, dim=-1) # [B]
loss = self.loss_fct(pos_score, neg_score)
else: # self.loss_type = 'CE'
test_item_emb = self.item_embedding.weight
logits = torch.matmul(seq_output, test_item_emb.transpose(0, 1))
loss = self.loss_fct(logits, pos_items)
reg_loss = self.reg_loss([self.item_embedding.weight,self.final_layer.weight])
loss = loss + self.reg_weight * reg_loss + self.reg_loss_rb()
return loss
[docs] def predict(self, interaction):
item_seq = interaction[self.ITEM_SEQ]
test_item = interaction[self.ITEM_ID]
seq_output = self.forward(item_seq)
test_item_emb = self.item_embedding(test_item)
scores = torch.mul(seq_output, test_item_emb).sum(dim=1)
return scores
[docs] def full_sort_predict(self, interaction):
item_seq = interaction[self.ITEM_SEQ]
# item_seq_len = interaction[self.ITEM_SEQ_LEN]
seq_output = self.forward(item_seq)
test_items_emb = self.item_embedding.weight
scores = torch.matmul(seq_output, test_items_emb.transpose(0, 1)) # [B, item_num]
return scores
[docs]class ResidualBlock_a(nn.Module):
r"""
Residual block (a) in the paper
"""
def __init__(self, in_channel, out_channel, kernel_size=3, dilation=None):
super(ResidualBlock_a, self).__init__()
half_channel = out_channel//2
self.ln1 = nn.LayerNorm(out_channel, eps=1e-8)
self.conv1 = nn.Conv2d(in_channel, half_channel, kernel_size=(1, 1), padding=0)
self.ln2 = nn.LayerNorm(half_channel, eps=1e-8)
self.conv2 = nn.Conv2d(half_channel, half_channel, kernel_size=(1, kernel_size), padding=0, dilation=dilation)
self.ln3 = nn.LayerNorm(half_channel, eps=1e-8)
self.conv3 = nn.Conv2d(half_channel, out_channel, kernel_size=(1, 1), padding=0)
self.dilation = dilation
self.kernel_size = kernel_size
[docs] def forward(self, x): # x: [batch_size, seq_len, embed_size]
out = F.relu(self.ln1(x))
out = out.permute(0, 2, 1).unsqueeze(2)
out = self.conv1(out).squeeze(2).permute(0, 2, 1)
out2 = F.relu(self.ln2(out))
out2 = self.conv_pad(out2, self.dilation)
out2 = self.conv2(out2).squeeze(2).permute(0, 2, 1)
out3 = F.relu(self.ln3(out2))
out3 = out3.permute(0, 2, 1).unsqueeze(2)
out3 = self.conv3(out3).squeeze(2).permute(0, 2, 1)
return out3 + x
[docs] def conv_pad(self, x, dilation): # x: [batch_size, seq_len, embed_size]
r""" Dropout-mask: To avoid the future information leakage problem, this paper proposed a masking-based dropout
trick for the 1D dilated convolution to prevent the network from seeing the future items.
Also the One-dimensional transformation is completed in this funtion.
"""
inputs_pad = x.permute(0, 2, 1) # [batch_size, embed_size, seq_len]
inputs_pad = inputs_pad.unsqueeze(2) # [batch_size, embed_size, 1, seq_len]
pad = nn.ZeroPad2d(((self.kernel_size - 1) * dilation, 0, 0, 0)) # padding opration args：(left,right,top,bottom)
inputs_pad = pad(inputs_pad) # [batch_size, embed_size, 1, seq_len+(self.kernel_size-1)*dilations]
return inputs_pad
[docs]class ResidualBlock_b(nn.Module):
r"""
Residual block (b) in the paper
"""
def __init__(self, in_channel, out_channel, kernel_size=3, dilation=None):
super(ResidualBlock_b, self).__init__()
self.conv1 = nn.Conv2d(in_channel, out_channel, kernel_size=(1, kernel_size), padding=0, dilation=dilation)
self.ln1 = nn.LayerNorm(out_channel, eps=1e-8)
self.conv2 = nn.Conv2d(out_channel, out_channel, kernel_size=(1, kernel_size), padding=0, dilation=dilation*2)
self.ln2 = nn.LayerNorm(out_channel, eps=1e-8)
self.dilation = dilation
self.kernel_size = kernel_size
[docs] def forward(self, x): # x: [batch_size, seq_len, embed_size]
x_pad = self.conv_pad(x, self.dilation) # [batch_size, embed_size, 1, seq_len+(self.kernel_size-1)*dilations]
out = self.conv1(x_pad).squeeze(2).permute(0, 2, 1) # [batch_size, seq_len+(self.kernel_size-1)*dilations-kernel_size+1, embed_size]
out = F.relu(self.ln1(out))
out_pad = self.conv_pad(out, self.dilation*2)
out2 = self.conv2(out_pad).squeeze(2).permute(0, 2, 1)
out2 = F.relu(self.ln2(out2))
return out2 + x
[docs] def conv_pad(self, x, dilation):
r""" Dropout-mask: To avoid the future information leakage problem, this paper proposed a masking-based dropout
trick for the 1D dilated convolution to prevent the network from seeing the future items.
Also the One-dimensional transformation is completed in this funtion.
"""
inputs_pad = x.permute(0, 2, 1)
inputs_pad = inputs_pad.unsqueeze(2)
pad = nn.ZeroPad2d(((self.kernel_size - 1) * dilation, 0, 0, 0))
inputs_pad = pad(inputs_pad)
return inputs_pad