NerualCF——CF与深度学习的结合

---

论文地址

Neural Collaborative Filtering

---

基本原理

将传统的矩阵分解中用户向量和物品向量的点积操作，换成由神经网络代替的互操作。

• 优点：更强表达能力的矩阵分解模型；
• 缺点：只使用了用户和物品的id特征，没有引入更多特征；

---

网络结构图

---

代码实现

import torch
import tqdm
import numpy as np
import pandas as pd
from torch import nn
from torch.utils import data
from collections import namedtuple
from sklearn.preprocessing import LabelEncoder
import argparse
from datetime import datetime

import warnings

warnings.filterwarnings("ignore")

# 使用具名元组定义特征标记
SparseFeat = namedtuple('SparseFeat', ['name', 'vocabulary_size', 'embedding_dim'])
DenseFeat = namedtuple('DenseFeat', ['name', 'dimension'])


# 数据集
class MovieLens(data.Dataset):
    def __init__(self, train_datas):
        self.train_datas = train_datas

    def __len__(self):
        return len(self.train_datas)

    def __getitem__(self, idx):
        return self.train_datas[idx]


# 残差模块
class Residual_block(nn.Module):
    def __init__(self, dim_stack, hidden_unit):
        super(Residual_block, self).__init__()
        self.linear1 = nn.Linear(dim_stack, hidden_unit)
        self.linear2 = nn.Linear(hidden_unit, dim_stack)
        self.relu = nn.ReLU()

    def forward(self, x):
        orig_x = x.clone()
        x = self.linear1(x)
        x = self.linear2(x)
        out = self.relu(x + orig_x)
        return out


# 模型架构
class NCF(nn.Module):
    def __init__(self,
                 embedding_classes,
                 embedding_dim=8,
                 hidden_unit=32):
        super(NCF, self).__init__()
        self.GMF_embedding = nn.ModuleList([nn.Embedding(ec + 1, embedding_dim) for ec in embedding_classes])
        self.MLP_embedding = nn.ModuleList([nn.Embedding(ec + 1, embedding_dim) for ec in embedding_classes])
        self.all_features_cat = embedding_dim * 2
        self.linear1 = nn.Linear(self.all_features_cat, hidden_unit)
        self.linear2 = nn.Linear(hidden_unit, self.all_features_cat)
        self.last_linear = nn.Linear(self.all_features_cat + embedding_dim, 1)
        self.relu = nn.ReLU()

    def forward(self, x):
        user_feature, movie_feature, rating = x[:, 0], x[:, 1], x[:, 2]

        GMF = torch.mul(self.GMF_embedding[0](user_feature), self.GMF_embedding[1](movie_feature))

        MLP = torch.cat((self.MLP_embedding[0](user_feature), self.MLP_embedding[1](movie_feature)), 1)
        MLP = self.linear1(MLP)
        MLP = self.linear2(MLP)
        MLP = self.relu(MLP)

        NeuMF = torch.cat((GMF, MLP), 1)
        out = self.last_linear(NeuMF)

        return {"predicts": out, "labels": rating}


# 获取到Embedding层的类别数
def cret_dataset_get_classes(data_root="../data/ml-1m/ratings.dat", batch_size=1, shuffle=True, num_workers=0):
    # 读取数据，NCF使用的特征只有user_id和item_id
    rnames = ['user_id', 'movie_id', 'rating', 'timestamp']
    data_df = pd.read_csv(data_root, sep='::', engine="python", names=rnames)

    lbe = LabelEncoder()
    data_df['user_id'] = lbe.fit_transform(data_df['user_id'])
    data_df['movie_id'] = lbe.fit_transform(data_df['movie_id'])

    train_data = data_df[['user_id', 'movie_id']]
    train_data['label'] = data_df['rating']

    dnn_feature_columns = [train_data['user_id'].nunique(), train_data['movie_id'].nunique()]

    train_dataset = MovieLens(train_data.to_numpy())
    train_loader = data.DataLoader(train_dataset, batch_size=batch_size, shuffle=shuffle, num_workers=num_workers)

    return train_loader, dnn_feature_columns


def train(config):
    train_loader, embedding_classes = cret_dataset_get_classes(config.data_root, config.batch_size)

    # 初始化模型
    model = NCF(embedding_classes,
                config.embedding_dim,
                hidden_unit=config.hidden_unit)

    # print(model)

    # 初始化损失函数
    loss_fn = nn.MSELoss()
    # 初始化优化器
    optimizer = torch.optim.Adam(model.parameters(), lr=config.lr)

    epoch = range(config.epoch)

    for epc in epoch:
        with tqdm.tqdm(
                iterable=train_loader,
                bar_format='{desc} {n_fmt:>4s}/{total_fmt:<4s} {percentage:3.0f}%|{bar}| {postfix}'
        ) as t:
            start_time = datetime.now()
            t.set_description_str(f"\33[36m【Epoch {epc + 1:04d}】")
            for batch in train_loader:
                out = model(batch)
                loss = loss_fn(out["predicts"].squeeze(1), out["labels"].float())
                # Backpropagation
                optimizer.zero_grad()
                loss.backward()
                optimizer.step()
                cur_time = datetime.now()
                delta_time = cur_time - start_time
                t.set_postfix_str(f"train_loss={loss.item():.7f}， 执行时长：{delta_time}\33[0m")
                t.update()


def test():
    pass


if __name__ == '__main__':
    parser = argparse.ArgumentParser(description='Process some integers.')
    parser.add_argument('--data_root',
                        default="../data/ml-1m/ratings.dat",
                        type=str,
                        help='an integer for the accumulator')
    parser.add_argument('--batch_size',
                        default=32,
                        type=int,
                        help='an integer for the accumulator')
    parser.add_argument('--lr',
                        default=1e-3,
                        type=float,
                        help='nothing')
    parser.add_argument('--epoch',
                        default=2,
                        type=int,
                        help='nothing')
    parser.add_argument('--embedding_dim',
                        default=8,
                        type=int,
                        help='nothing')
    parser.add_argument('--hidden_unit',
                        default=64,
                        type=int,
                        help='nothing')
    config = parser.parse_args()
    train(config)

---

总结分析

• Embedding

首先还是对输入的数据分别进行Embedding操作，需要构建两个Embedding操作块；

• GMF Layer

在这里是针对MovieLens中的共现矩阵组成部分，“用户”、“电影”两个特征首先进行Embedding后，进行内积；

• MLP Layer

首先对“用户”、“电影”两个特征进行Concat，之后送入MLP层推理；

• NeuMF Layer

将GMF和MLP的结果进行Concat，再经过一个Linear层输出一维的结果；

• Loss

用MSE损失计算和GT间的损失值。

• 问题

训练过程不稳定，难以收敛？

---

参考文献

Neural Collaborative Filtering
【翻译】Neural Collaborative Filtering–神经协同过滤
相关实现

---