Skip to content

xDeepFM

潜心 edited this page Oct 8, 2020 · 3 revisions

xDeepFM

1. 论文

xDeepFM: Combining Explicit and Implicit Feature Interactions for Recommender Systems

创新:CIN层是关键!!!

原文笔记: https://mp.weixin.qq.com/s/TohOmVpQzNlA3vXv0gpobg

2. 模型结构

3. 实验数据集

采用Criteo数据集进行测试。数据集的处理见utils文件,主要分为:

  1. 考虑到Criteo文件过大,因此可以通过read_partsample_sum读取部分数据进行测试;
  2. 对缺失数据进行填充;
  3. 对密集数据I1-I13进行归一化处理,对稀疏数据C1-C26进行重新编码LabelEncoder
  4. 整理得到feature_columns
  5. 切分数据集,最后返回feature_columns, (train_X, train_y), (test_X, test_y)

4. 模型API

class xDeepFM(keras.Model):
    def __init__(self, feature_columns, hidden_units, cin_size, dnn_dropout=0, dnn_activation='relu',
                 embed_reg=1e-5, cin_reg=1e-5):
        """
        xDeepFM
        :param feature_columns: A list. a list containing dense and sparse column feature information.
        :param hidden_units: A list. a list of dnn hidden units.
        :param cin_size: A list. a list of the number of CIN layers.
        :param dnn_dropout: A scalar. dropout of dnn.
        :param dnn_activation: A string. activation function of dnn.
        :param embed_reg: A scalar. the regularizer of embedding.
        :param cin_reg: A scalar. the regularizer of cin.
        """

5. 实验超参数

  • file:Criteo文件;
  • read_part:是否读取部分数据,True
  • sample_num:读取部分时,样本数量,5000000
  • test_size:测试集比例,0.2
  • embed_dim:Embedding维度,8
  • dnn_dropout:Dropout, 0.5
  • hidden_unit:DNN的隐藏单元,[256, 128, 64]
  • cin_size:CIN尺度,(128, 128)
  • learning_rate:学习率,0.001
  • batch_size:4096
  • epoch:10

6. 实验结果

采用Criteo数据集中前500w条数据,最终测试集的结果为:AUC:0.738484

Experimental dataset

Model

Clone this wiki locally