# -*- coding: utf-8 -*-
# @Time : 2020/7/8
# @Author : Shanlei Mu
# @Email : slmu@ruc.edu.cn
# @File : deepfm.py
# UPDATE:
# @Time : 2020/8/14
# @Author : Zihan Lin
# @Email : linzihan.super@foxmain.com
r"""
DeepFM
################################################
Reference:
Huifeng Guo et al. "DeepFM: A Factorization-Machine based Neural Network for CTR Prediction." in IJCAI 2017.
"""
import torch.nn as nn
from torch.nn.init import xavier_normal_, constant_
from recbole.model.abstract_recommender import ContextRecommender
from recbole.model.layers import BaseFactorizationMachine, MLPLayers
[docs]class DeepFM(ContextRecommender):
"""DeepFM is a DNN enhanced FM which both use a DNN and a FM to calculate feature interaction.
Also DeepFM can be seen as a combination of FNN and FM.
"""
def __init__(self, config, dataset):
super(DeepFM, self).__init__(config, dataset)
# load parameters info
self.mlp_hidden_size = config["mlp_hidden_size"]
self.dropout_prob = config["dropout_prob"]
# define layers and loss
self.fm = BaseFactorizationMachine(reduce_sum=True)
size_list = [
self.embedding_size * self.num_feature_field
] + self.mlp_hidden_size
self.mlp_layers = MLPLayers(size_list, self.dropout_prob)
self.deep_predict_layer = nn.Linear(
self.mlp_hidden_size[-1], 1
) # Linear product to the final score
self.sigmoid = nn.Sigmoid()
self.loss = nn.BCEWithLogitsLoss()
# parameters initialization
self.apply(self._init_weights)
def _init_weights(self, module):
if isinstance(module, nn.Embedding):
xavier_normal_(module.weight.data)
elif isinstance(module, nn.Linear):
xavier_normal_(module.weight.data)
if module.bias is not None:
constant_(module.bias.data, 0)
[docs] def forward(self, interaction):
deepfm_all_embeddings = self.concat_embed_input_fields(
interaction
) # [batch_size, num_field, embed_dim]
batch_size = deepfm_all_embeddings.shape[0]
y_fm = self.first_order_linear(interaction) + self.fm(deepfm_all_embeddings)
y_deep = self.deep_predict_layer(
self.mlp_layers(deepfm_all_embeddings.view(batch_size, -1))
)
y = y_fm + y_deep
return y.squeeze(-1)
[docs] def calculate_loss(self, interaction):
label = interaction[self.LABEL]
output = self.forward(interaction)
return self.loss(output, label)
[docs] def predict(self, interaction):
return self.sigmoid(self.forward(interaction))