关注我们，一起学习~
之前和大家分享了动手实现wide and deep和deepfm，今天和大家分享动手实现DCN。和之前连个方法类似，该方法也是考虑记忆和泛化两部分，泛化部分同样采用DNNs；而记忆部分采用了不同的方式，wide & deep和deepfm我们可以理解为都进行了一次交互，而DCN进行多次交互，每次进行特征的显式交叉，经过一层全连接层后加上原始输入（残差的思想），具体如下图所示。不过有个缺点是因为cross interaction之后得到的是将1n的向量变成nn的矩阵，所以占用内存会增大很多，需要考虑向量的大小。
具体代码如下，采用pytorch实现。
代码语言：javascript复制'''
Author: dqdallen 1986360994@qq.com
Date: 2022-06-15 10:38:25
LastEditors: dqdallen 1986360994@qq.com
LastEditTime: 2022-06-30 09:26:05
FilePath: recommendationDCNdeep_cross_network.py
Description: 这是默认设置,请设置`customMade`, 打开koroFileHeader查看配置 进行设置: https://github.com/OBKoro1/koro1FileHeader/wiki/配置
'''
import torch
import torch.nn as nn
class DCN(nn.Module):
def __init__(self, hidden_dims, cross_num, emb_dim, dropouts,
fea_nums_dict, numb_dim):
"""_summary_
Args:
hidden_dims (list): the dims of each hidden layer
cross_num (int): the number of features cross interaction
emb_dim (int): the dim of embedding
dropouts (list): the dropout rates
fea_nums_dict (dict): the number of each category feature
numb_dim (int): the number of number feature
"""
super(DCN, self).__init__()
self.hidden_dims = hidden_dims
self.cross_num = cross_num
self.emb_dim = emb_dim
self.dropouts = dropouts [0]
self.fea_nums_dict = fea_nums_dict
# 不同特征的embedding层
self.emb_layer = nn.ModuleDict()
for fea_name in self.fea_nums_dict.keys():
self.emb_layer[fea_name] = nn.Embedding(
self.fea_nums_dict[fea_name], emb_dim)
# DNNs部分
input_dim = len(self.fea_nums_dict.values()) * emb_dim numb_dim
hidden_dims = [input_dim] hidden_dims
self.dnns = nn.ModuleList([nn.BatchNorm1d(input_dim)])
for i in range(len(hidden_dims) - 1):
self.dnns.append(nn.Linear(hidden_dims[i], hidden_dims[i 1]))
self.dnns.append(nn.BatchNorm1d(hidden_dims[i 1]))
self.dnns.append(nn.Dropout(self.dropouts[i]))
# 特征交叉交互部分
self.corss_layer = nn.ModuleList()
for _ in range(cross_num):
self.corss_layer.append(nn.Linear(input_dim - numb_dim, 1))
self.predict = nn.Linear(
hidden_dims[-1] emb_dim * len(self.fea_nums_dict.values()), 1)
self.sigmoid = nn.Sigmoid()
def coss_layer(self, features):
# 获得embedding
emb_arr = []
for fea_name in features:
emb_tmp = self.emb_layer[fea_name](features[fea_name])
emb_arr.append(emb_tmp)
self.embedding = torch.cat(emb_arr, 1)
emb = self.embedding
# 特征进行cross interaction
for i in range(self.cross_num):
emb_tmp = torch.bmm(torch.transpose(emb.unsqueeze(1), 1, 2),
emb.unsqueeze(1))
emb_tmp = self.corss_layer[i](emb_tmp)
emb = emb_tmp.transpose(1, 2).squeeze(1) emb
return emb
def deep_layer(self, numb_features):
deep_emb = torch.cat([self.embedding, numb_features], 1)
for layer in self.dnns:
deep_emb = layer(deep_emb)
return deep_emb
def forward(self, numb_features, features):
cross_result = self.coss_layer(features)
dnn_result = self.deep_layer(numb_features)
output = torch.cat([cross_result, dnn_result], 1)
output = self.predict(output)
output = self.sigmoid(output)
return output
