关注我们，一起学习~
基本原理
DeepFM方法是华为在2017年提出的，是推荐系统中的经典方法，和wide and deep类似，DeepFM也是同时考虑记忆和泛化，在泛化部分采用DNNs结合dropout，batch norm提升模型泛化能力；而记忆部分，采用的是FM，将特征之间进行两两交互。最后，将DNNs部分和FM部分的输出拼接，对偏好分数进行预测。原理相对简单，这里不做过多介绍。
想要详细了解的小伙伴可以前往链接https://www.jianshu.com/p/4ec8d8dcb660
代码实现
代码主要是基于pytorch实现，给出了对应的注释，希望对大家有一点点帮助。
代码语言：javascript复制import torch
import torch.nn as nn
class DeepFM(nn.Module):
def __init__(self, hidden_dims, emb_dim, dropouts, fea_nums_dict, numb_dim):
"""deepfm模型
Args:
hidden_dims (list): the hidden dims of dnn
emb_dim (int): the dim of embedding
dropouts (list): the rates of dropouts in dnn
fea_nums_dict (dict): each category's feature nums
numb_dim (int): the dim of numerical features
"""
super(DeepFM, self).__init__()
self.hidden_dims = hidden_dims
self.emb_dim = emb_dim
self.dropouts = dropouts [0]
self.fea_nums_dict = fea_nums_dict
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
)
self.fm_first = nn.ModuleDict()
for fea_name in self.fea_nums_dict.keys():
self.fm_first[fea_name] = nn.Embedding(
self.fea_nums_dict[fea_name], 1)
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]))
self.predict = nn.Linear(
hidden_dims[-1] emb_dim len(self.fea_nums_dict.values()), 1
)
self.sigmoid = nn.Sigmoid()
def fm_layer(self, features):
# 一阶交互
emb_arr = []
for fea_name in features:
emb_tmp = self.fm_first[fea_name](features[fea_name])
emb_arr.append(emb_tmp)
first_order = torch.cat(emb_arr, 1)
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)
# 二阶交互
sum_square = sum(emb_arr) ** 2
square_sum = sum([emb * emb for emb in emb_arr])
second_order = 0.5 * (sum_square - square_sum)
# 拼接一阶和二阶交互
fm_result = torch.cat([first_order, second_order], 1)
return fm_result
def deep_layer(self, numb_features):
"""dnns层
Args:
numb_features (tensor): numerical features
Returns:
tensor: dnn_result
"""
# 拼接类别特征的embedding和数值特征
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):
fm_result = self.fm_layer(features)
dnn_result = self.deep_layer(numb_features)
output = torch.cat([fm_result, dnn_result], 1)
output = self.predict(output)
output = self.sigmoid(output)
return output
