# -*- coding: utf-8 -*-
# @Time : 2020/7/21
# @Author : Zihan Lin
# @Email : linzihan.super@foxmail.com
# @File : afm.py
r"""
AFM
################################################
Reference:
Jun Xiao et al. "Attentional Factorization Machines: Learning the Weight of Feature Interactions via
Attention Networks" in IJCAI 2017.
"""
import torch
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 AttLayer
[docs]class AFM(ContextRecommender):
""" AFM is a attention based FM model that predict the final score with the attention of input feature.
"""
def __init__(self, config, dataset):
super(AFM, self).__init__(config, dataset)
# load parameters info
self.attention_size = config['attention_size']
self.dropout_prob = config['dropout_prob']
self.reg_weight = config['reg_weight']
self.num_pair = self.num_feature_field * (self.num_feature_field - 1) / 2
# define layers and loss
self.attlayer = AttLayer(self.embedding_size, self.attention_size)
self.p = nn.Parameter(torch.randn(self.embedding_size), requires_grad=True)
self.dropout_layer = nn.Dropout(p=self.dropout_prob)
self.sigmoid = nn.Sigmoid()
self.loss = nn.BCELoss()
# 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 build_cross(self, feat_emb):
""" Build the cross feature columns of feature columns
Args:
feat_emb (torch.FloatTensor): input feature embedding tensor. shape of [batch_size, field_size, embed_dim].
Returns:
tuple:
- torch.FloatTensor: Left part of the cross feature. shape of [batch_size, num_pairs, emb_dim].
- torch.FloatTensor: Right part of the cross feature. shape of [batch_size, num_pairs, emb_dim].
"""
# num_pairs = num_feature_field * (num_feature_field-1) / 2
row = []
col = []
for i in range(self.num_feature_field - 1):
for j in range(i + 1, self.num_feature_field):
row.append(i)
col.append(j)
p = feat_emb[:, row] # [batch_size, num_pairs, emb_dim]
q = feat_emb[:, col] # [batch_size, num_pairs, emb_dim]
return p, q
[docs] def afm_layer(self, infeature):
""" Get the attention-based feature interaction score
Args:
infeature (torch.FloatTensor): input feature embedding tensor. shape of [batch_size, field_size, embed_dim].
Returns:
torch.FloatTensor: Result of score. shape of [batch_size, 1].
"""
p, q = self.build_cross(infeature)
pair_wise_inter = torch.mul(p, q) # [batch_size, num_pairs, emb_dim]
# [batch_size, num_pairs, 1]
att_signal = self.attlayer(pair_wise_inter).unsqueeze(dim=2)
att_inter = torch.mul(att_signal, pair_wise_inter) # [batch_size, num_pairs, emb_dim]
att_pooling = torch.sum(att_inter, dim=1) # [batch_size, emb_dim]
att_pooling = self.dropout_layer(att_pooling) # [batch_size, emb_dim]
att_pooling = torch.mul(att_pooling, self.p) # [batch_size, emb_dim]
att_pooling = torch.sum(att_pooling, dim=1, keepdim=True) # [batch_size, 1]
return att_pooling
[docs] def forward(self, interaction):
afm_all_embeddings = self.concat_embed_input_fields(interaction) # [batch_size, num_field, embed_dim]
output = self.sigmoid(self.first_order_linear(interaction) + self.afm_layer(afm_all_embeddings))
return output.squeeze(-1)
[docs] def calculate_loss(self, interaction):
label = interaction[self.LABEL]
output = self.forward(interaction)
l2_loss = self.reg_weight * torch.norm(self.attlayer.w.weight, p=2)
return self.loss(output, label) + l2_loss
[docs] def predict(self, interaction):
return self.forward(interaction)