DeepCTR项目常见问题解答与技术指南

最新推荐文章于 2025-06-17 02:43:20 发布

江奎钰

最新推荐文章于 2025-06-17 02:43:20 发布

阅读量282

点赞数 4

CC 4.0 BY-SA版权

本文链接：https://blog.youkuaiyun.com/gitblog_00973/article/details/148443007

DeepCTR项目常见问题解答与技术指南

DeepCTR Easy-to-use,Modular and Extendible package of deep-learning based CTR models . 项目地址: https://gitcode.com/gh_mirrors/de/DeepCTR

前言

DeepCTR是一个基于深度学习的点击率预测工具库，它集成了多种先进的深度CTR模型，如DeepFM、AFM等。本文将针对使用DeepCTR过程中常见的技术问题，提供详细的解决方案和最佳实践指南。

模型保存与加载

在DeepCTR中，模型的保存与加载遵循Keras的标准方式，但需要注意一些特殊处理：

权重保存与加载

# 创建DeepFM模型
model = DeepFM()

# 保存权重
model.save_weights('DeepFM_w.h5')

# 加载权重
model.load_weights('DeepFM_w.h5')

完整模型保存与加载

完整模型保存需要特别注意自定义层的处理：

from tensorflow.python.keras.models import save_model, load_model
from deepctr.layers import custom_objects

# 保存完整模型
model = DeepFM()
save_model(model, 'DeepFM.h5')

# 加载完整模型时需要传入custom_objects参数
model = load_model('DeepFM.h5', custom_objects)

学习率设置与早停机制

Keras风格模型设置

from tensorflow.python.keras.optimizers import Adam, Adagrad
from tensorflow.python.keras.callbacks import EarlyStopping

# 创建模型并设置优化器
model = deepctr.models.DeepFM(linear_feature_columns, dnn_feature_columns)
model.compile(Adagrad(0.1024), 'binary_crossentropy', metrics=['binary_crossentropy'])

# 设置早停机制
es = EarlyStopping(monitor='val_binary_crossentropy')
history = model.fit(model_input, data[target].values, 
                   batch_size=256, epochs=10, verbose=2, 
                   validation_split=0.2, callbacks=[es])

Estimator风格模型设置

import tensorflow as tf
from deepctr.estimator import DeepFMEstimator

# 分别设置线性部分和DNN部分的优化器
model = DeepFMEstimator(linear_feature_columns, dnn_feature_columns, task='binary',
                       linear_optimizer=tf.train.FtrlOptimizer(0.05),
                       dnn_optimizer=tf.train.AdagradOptimizer(0.1))

AFM模型注意力权重获取

AFM(Attention Factorization Machine)模型中的特征交互注意力权重可以按以下方式提取：

import itertools
from tensorflow.python.keras.models import Model
from tensorflow.python.keras.layers import Lambda

# 训练AFM模型
model = AFM(linear_feature_columns, dnn_feature_columns)
model.fit(model_input, target)

# 构建注意力权重提取模型
afmlayer = model.layers[-3]
afm_weight_model = Model(model.input, 
                        outputs=Lambda(lambda x: afmlayer.normalized_att_score)(model.input))

# 获取所有样本的注意力权重
attentional_weights = afm_weight_model.predict(model_input, batch_size=4096)

# 获取特征交互对
feature_names = get_feature_names(dnn_feature_columns)
feature_interactions = list(itertools.combinations(feature_names, 2))

嵌入向量提取方法

从DeepFM等模型中提取嵌入向量：

feature_columns = [SparseFeat('user_id',120), SparseFeat('item_id',60), SparseFeat('cate_id',60)]

def get_embedding_weights(dnn_feature_columns, model):
    embedding_dict = {}
    for fc in dnn_feature_columns:
        if hasattr(fc, 'embedding_name'):
            name = fc.embedding_name if fc.embedding_name is not None else fc.name
            embedding_dict[name] = model.get_layer("sparse_emb_"+name).get_weights()[0]
    return embedding_dict
    
# 获取嵌入向量字典
embedding_dict = get_embedding_weights(feature_columns, model)

# 提取特定特征的嵌入向量
user_id_emb = embedding_dict['user_id']
item_id_emb = embedding_dict['item_id']

处理密集特征输入

在模型中添加密集特征向量输入：

from deepctr.feature_column import SparseFeat, DenseFeat
import numpy as np

# 定义特征列，包含稀疏特征和密集特征
feature_columns = [SparseFeat('user_id',120), 
                  SparseFeat('item_id',60),
                  DenseFeat("pic_vec",5)]

# 准备输入数据
user_id = np.array([[1],[0],[1]])
item_id = np.array([[30],[20],[10]])
pic_vec = np.array([[0.1,0.5,0.4,0.3,0.2],
                   [0.1,0.5,0.4,0.3,0.2],
                   [0.1,0.5,0.4,0.3,0.2]])
label = np.array([1,0,1])

model_input = {'user_id':user_id, 'item_id':item_id, 'pic_vec':pic_vec}

# 构建并训练模型
model = DeepFM(feature_columns, feature_columns)
model.compile('adagrad', 'binary_crossentropy')
model.fit(model_input, label)

预训练嵌入权重初始化与冻结

使用预训练权重初始化嵌入层并冻结训练：

import tensorflow as tf

# 准备预训练权重
pretrained_item_weights = np.random.randn(60,4)
pretrained_weights_initializer = tf.initializers.constant(pretrained_item_weights)

# 定义特征列，设置初始化器和trainable=False
feature_columns = [SparseFeat('user_id',120),
                  SparseFeat('item_id',60, embedding_dim=4,
                           embeddings_initializer=pretrained_weights_initializer,
                           trainable=False)]

# 准备输入数据
user_id = np.array([[1],[0],[1]])
item_id = np.array([[30],[20],[10]])
label = np.array([1,0,1])

model_input = {'user_id':user_id, 'item_id':item_id}

# 构建并训练模型
model = DeepFM(feature_columns, feature_columns)
model.compile('adagrad', 'binary_crossentropy')
model.fit(model_input, label)