点击量预测(CTR)——Wide&Deep理论与实践

一、Wide&Deep模型

谷歌于2016年提出的Wide&Deep模型。Wide&Deep模型的主要思路正如其名，是由单层的Wide部分和多层的Deep部分组成的混合模型。其中Wide部分让模型拥有“记忆能力（memorization）”；Deep部分的主要作用是让模型容易“泛化能力（generalization）”。

Wide&Deep模型的paper地址如下： https://arxiv.org/pdf/1606.07792v1.pdf
Wide&Deep模型主要特点： 使模型同时具有逻辑回归和深度神经网络的优点——能够快速处理并记忆大量历史行为特征，并且具有强大的表达能力。
Wide&Deep模型的优点：
* 简单有效。结构简单易于理解，效果优异。目前仍在工业界广泛使用，也证明了该模型的有效性。
* 结构新颖。使用不同于以往的线性模型与DNN串行连接的方式，而将线性模型与DNN并行连接，同时兼顾模型的Memorization与Generalization。

Wide&Deep模型的缺点：
* Wide侧的特征工程仍无法避免。

二、算法原理

2.1 Wide部分——“记忆能力”

wide部分是一个广义的线性模型，可以参考逻辑回归模型。

“记忆能力”： 可以被理解为模型的直接学习并利用历史数据中物品或者特征的“共显频率”的能力。

像以前学习的逻辑回归、协同过滤等简单的模型有较强的“记忆能力”。这类模型类似于“如果点击A,就推荐B”这类似的规则式的推荐，这就相当于模型直接记住了历史数据的分布特点。

2.2 Deep部分——“泛化能力”

Deep部分是一个经典的DNN模型。

“泛化能力” 可以被理解为模型的传递特征的相关性，以及挖掘稀疏甚至从未出现过的稀有特征与最终的标签相关性能力。

像矩阵分解比协同过滤的泛化能力强，由于矩阵分解引入了隐向量的结构；
像DNN模型，通过特征多次的自由组合，可以深度挖掘数据中的潜在的模式；这两种使得稀疏的数据，获得有数据支撑，使得模型能得到稳定的推荐。这就是简单模型所缺乏的“泛化能力”。

三、代码实现

采取的数据是movielens 100.为了操作的方便，只为了展示FM实现的过程，只选取了uid、itemId作为输入特征，rating作为lable。

数据集

u.item： 电影信息数据

 movie id | movie title | release date | video release date |IMDb URL |unknown | Action | Adventure | Animation |Children's | Comedy | Crime |Documentary | Drama | Fantasy |Film-Noir | Horror | Musical | Mystery |Romance | Sci-Fi |Thriller | War | Western
 
1|Toy Story (1995)|01-Jan-1995||http://us.imdb.com/M/title-exact?Toy%20Story%20(1995)|0|0|0|1|1|1|0|0|0|0|0|0|0|0|0|0|0|0|0
2|GoldenEye (1995)|01-Jan-1995||http://us.imdb.com/M/title-exact?GoldenEye%20(1995)|0|1|1|0|0|0|0|0|0|0|0|0|0|0|0|0|1|0|0
3|Four Rooms (1995)|01-Jan-1995||http://us.imdb.com/M/title-exact?Four%20Rooms%20(1995)|0|0|0|0|0|0|0|0|0|0|0|0|0|0|0|0|1|0|0

u.user： 用户信息数据

user id | age | gender | occupation | zip code

1|24|M|technician|85711
2|53|F|other|94043
3|23|M|writer|32067

ua.base： 训练数据集
ua.test: 测试数据集

user id | item id | rating | timestamp
1	1	5	874965758
1	2	3	876893171
1	3	4	878542960

数据处理

将uid和itemId使用one-hot编码，将rating作为输出标签，其评分等级为[0-5]，大于3为1（表示用户感兴趣）小于3为0（表示用户不感兴趣）。

# 数据加载
def loadData():
    # user信息(只取uid)
    userInfo = pd.read_csv('../data/u.user', sep='\|', names=['uid', 'age', 'gender', 'occupation','zip code'])
    uid_ = userInfo['uid']
    userId_dum = pd.get_dummies(userInfo['uid'], columns=['uid'], prefix='uid_')
    userId_dum['uid']=uid_

    # item信息（只取itemId）
    header = ['item_id', 'title', 'release_date', 'video_release_date', 'IMDb_URL', 'unknown', 'Action', 'Adventure', 'Animation', 'Children',
              'Comedy', 'Crime', 'Documentary', 'Drama', 'Fantasy', 'Film-Noir', 'Horror', 'Musical', 'Mystery', 'Romance', 'Sci-Fi',
              'Thriller', 'War', 'Western']
    ItemInfo = pd.read_csv('../data/u.item', sep='|', names=header, encoding = "ISO-8859-1")
    ItemInfo = ItemInfo.drop(columns=['title', 'release_date', 'video_release_date', 'IMDb_URL', 'unknown'])
    item_id_ = ItemInfo['item_id']
    item_Id_dum = pd.get_dummies(ItemInfo['item_id'], columns=['item_id'], prefix='item_id_')
    item_Id_dum['item_id']=item_id_

    # 训练数据
    trainData = pd.read_csv('../data/ua.base', sep='\t', names=['uid', 'item_id', 'rating', 'time'])
    trainData = trainData.drop(columns=['time'])

    trainData['rating']=trainData.rating.apply(lambda x:1 if int(x)>3 else 0)

    Y_train=pd.get_dummies(trainData['rating'],columns=['rating'],prefix='y_')

    X_train = pd.merge(trainData, userId_dum, how='left')
    X_train = pd.merge(X_train, item_Id_dum, how='left')
    X_train=X_train.drop(columns=['uid','item_id','rating'])


    # 测试数据
    testData = pd.read_csv('../data/ua.test', sep='\t', names=['uid', 'item_id', 'rating', 'time'])
    testData = testData.drop(columns=['time'])

    testData['rating']=testData.rating.apply(lambda x:1 if int(x)>3 else 0)
    Y_test=pd.get_dummies(testData['rating'],columns=['rating'],prefix='y_')

    X_test = pd.merge(testData, userId_dum, how='left')
    X_test = pd.merge(X_test, item_Id_dum, how='left')
    X_test=X_test.drop(columns=['uid','item_id','rating'])

    # 对应域 uid itemid
    # user信息 uid
    # item信息 itemid
    field_index={
    
    }
    userField=['uid']
    itemField=['itemId']
    field=userField+itemField

    # 每个域的长度
    userFieldLen=[len(uid_)]
    itemFieldLen=[len(item_id_)]

    field_len = userFieldLen + itemFieldLen
    j=0
    field_arange=[0]
    for field_n in  range(len(field)):
        field_arange.append(field_arange[field_n]+field_len[field_n])
    return X_train.values,Y_train.values,X_test.values,Y_test.values,field_arange,len(field)

Wide&Deep模型

class WideOrDeep:
    def __init__(self,vec_dim,linear_lr,dnn_lr,l1_reg ,feature_length,field_arange,field_len,dnn_layers):
        self.vec_dim=vec_dim
        self.linear_lr=linear_lr
        self.dnn_lr = dnn_lr
        self.l1_reg  = l1_reg
        self.feature_length=feature_length
        self.field_arange=field_arange
        self.field_len=field_len
        self.dnn_layers=dnn_layers

    def add_input(self):
        self.X = tf.placeholder(tf.float32,name='input_x',shape=[None,self.feature_length])
        self.Y = tf.placeholder(tf.float32, shape=[None,2], name='input_y')

    # 创建计算规则
    def inference(self):

        with tf.variable_scope('linear_part'):
            b = tf.get_variable(name='b', shape=[2], dtype=tf.float32)
            linear_w = tf.get_variable(shape=[self.feature_length, 2], dtype=tf.float32,
                                       name='linear_w')
            self.wide = tf.matmul(self.X, linear_w)+b


        with tf.variable_scope('dnn_layer'):
            Embedding = [tf.get_variable(name='Embedding_%d'%i, shape=[self.field_arange[i+1]-self.field_arange[i], self.vec_dim], dtype=tf.float32) for i in range(self.field_len)]
            Embedding_layer = tf.concat([tf.matmul(tf.slice(self.X,[0,self.field_arange[i]],[-1,self.field_arange[i+1]-self.field_arange[i]]), Embedding[i]) for i in range(self.field_len)], axis=1)
        x = Embedding_layer

        in_num = self.field_len * self.vec_dim

        for i in range(len(self.dnn_layers)):
            out_num = self.dnn_layers[i]
            w = tf.get_variable(name='w_%d'%i, shape=[in_num, out_num], dtype=tf.float32)
            b = tf.get_variable(name='b_%d'%i, shape=[out_num], dtype=tf.float32)
            x = tf.matmul(x, w) + b
            if out_num == 2:
                self.y_out = x + self.wide+b
            else:
                x = tf.nn.relu(x)
            in_num = out_num
    def add_loss(self):
        self.loss = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(labels=self.Y, logits=self.y_out))

    #计算accuracy
    def add_accuracy(self):
        # accuracy
        self.correct_prediction = tf.equal(tf.cast(tf.argmax(self.y_out,1), tf.float32), tf.cast(tf.argmax(self.Y,1), tf.float32))
        self.accuracy = tf.reduce_mean(tf.cast(self.correct_prediction, tf.float32))

        # self.auc_value = tf.metrics.auc(tf.argmax(self.y_out,1),tf.argmax(self.Y,1), curve='ROC')
    #训练
    def train(self):

        wide_opt = tf.train.FtrlOptimizer(learning_rate=self.linear_lr, l1_regularization_strength=self.l1_reg)
        wide_opt_train = wide_opt.minimize(loss=self.loss)

        deep_opt = tf.train.AdamOptimizer(learning_rate=self.dnn_lr)
        deep_opt_train = deep_opt.minimize(loss=self.loss)

        self.train_op = tf.group(wide_opt_train, deep_opt_train)

    #构建图
    def build_graph(self):
        self.add_input()
        self.inference()
        self.add_loss()
        self.add_accuracy()
        self.train()

训练和测试

def train_model(sess, model, X_train,Y_train,batch_size, epochs=100):
    num = len(X_train) // batch_size+1
    for step in range(epochs):
        print("epochs{0}:".format(step+1))
        for i in range(num):
            index = np.random.choice(len(X_train), batch_size)
            batch_x = X_train[index]
            batch_y = Y_train[index]
            feed_dict = {
    
    model.X: batch_x,
                         model.Y: batch_y}
            sess.run(model.train_op, feed_dict=feed_dict)


            # print("Iteration {0}: with minibatch  training loss = {1}"
            #       .format(step+1, loss))

            if (i+1)%100==0:
                loss ,accuracy,y_out= sess.run([model.loss,model.accuracy,model.y_out], feed_dict=feed_dict)
                auc = metrics.roc_auc_score(batch_y, y_out)
                print("Iteration {0}: with minibatch training loss = {1} accuracy = {2} auc={3}"
                      .format(step+1, loss,accuracy,auc))

def test_model(sess,model,X_test,Y_test):

    loss,y_out, accuracy= sess.run([model.loss, model.y_out,model.accuracy], feed_dict={
    
    model.X: X_test, model.Y: Y_test})

    print("loss={0} accuracy={1} auc={2}".format(loss,accuracy,metrics.roc_auc_score(Y_test, y_out)))

完整代码

import tensorflow as tf
import numpy as np
import pandas as pd
from sklearn import metrics


def loadData():
    # user信息(只取uid)
    userInfo = pd.read_csv('../data/u.user', sep='\|', names=['uid', 'age', 'gender', 'occupation','zip code'])
    uid_ = userInfo['uid']
    userId_dum = pd.get_dummies(userInfo['uid'], columns=['uid'], prefix='uid_')
    userId_dum['uid']=uid_

    # item信息（只取itemId）
    header = ['item_id', 'title', 'release_date', 'video_release_date', 'IMDb_URL', 'unknown', 'Action', 'Adventure', 'Animation', 'Children',
              'Comedy', 'Crime', 'Documentary', 'Drama', 'Fantasy', 'Film-Noir', 'Horror', 'Musical', 'Mystery', 'Romance', 'Sci-Fi',
              'Thriller', 'War', 'Western']
    ItemInfo = pd.read_csv('../data/u.item', sep='|', names=header, encoding = "ISO-8859-1")
    ItemInfo = ItemInfo.drop(columns=['title', 'release_date', 'video_release_date', 'IMDb_URL', 'unknown'])
    item_id_ = ItemInfo['item_id']
    item_Id_dum = pd.get_dummies(ItemInfo['item_id'], columns=['item_id'], prefix='item_id_')
    item_Id_dum['item_id']=item_id_

    # 训练数据
    trainData = pd.read_csv('../data/ua.base', sep='\t', names=['uid', 'item_id', 'rating', 'time'])
    trainData = trainData.drop(columns=['time'])

    trainData['rating']=trainData.rating.apply(lambda x:1 if int(x)>3 else 0)

    Y_train=pd.get_dummies(trainData['rating'],columns=['rating'],prefix='y_')

    X_train = pd.merge(trainData, userId_dum, how='left')
    X_train = pd.merge(X_train, item_Id_dum, how='left')
    X_train=X_train.drop(columns=['uid','item_id','rating'])


    # 测试数据
    testData = pd.read_csv('../data/ua.test', sep='\t', names=['uid', 'item_id', 'rating', 'time'])
    testData = testData.drop(columns=['time'])

    testData['rating']=testData.rating.apply(lambda x:1 if int(x)>3 else 0)
    Y_test=pd.get_dummies(testData['rating'],columns=['rating'],prefix='y_')

    X_test = pd.merge(testData, userId_dum, how='left')
    X_test = pd.merge(X_test, item_Id_dum, how='left')
    X_test=X_test.drop(columns=['uid','item_id','rating'])

    # 对应域 uid itemid
    # user信息 uid
    # item信息 itemid
    field_index={
    
    }
    userField=['uid']
    itemField=['itemId']
    field=userField+itemField

    # 每个域的长度
    userFieldLen=[len(uid_)]
    itemFieldLen=[len(item_id_)]

    field_len = userFieldLen + itemFieldLen
    j=0
    field_arange=[0]
    for field_n in  range(len(field)):
        field_arange.append(field_arange[field_n]+field_len[field_n])
    return X_train.values,Y_train.values,X_test.values,Y_test.values,field_arange,len(field)


class WideOrDeep:
    def __init__(self,vec_dim,linear_lr,dnn_lr,l1_reg ,feature_length,field_arange,field_len,dnn_layers):
        self.vec_dim=vec_dim
        self.linear_lr=linear_lr
        self.dnn_lr = dnn_lr
        self.l1_reg  = l1_reg
        self.feature_length=feature_length
        self.field_arange=field_arange
        self.field_len=field_len
        self.dnn_layers=dnn_layers

    def add_input(self):
        self.X = tf.placeholder(tf.float32,name='input_x',shape=[None,self.feature_length])
        self.Y = tf.placeholder(tf.float32, shape=[None,2], name='input_y')

    # 创建计算规则
    def inference(self):

        with tf.variable_scope('linear_part'):
            b = tf.get_variable(name='b', shape=[2], dtype=tf.float32)
            linear_w = tf.get_variable(shape=[self.feature_length, 2], dtype=tf.float32,
                                       name='linear_w')
            self.wide = tf.matmul(self.X, linear_w)+b


        with tf.variable_scope('dnn_layer'):
            Embedding = [tf.get_variable(name='Embedding_%d'%i, shape=[self.field_arange[i+1]-self.field_arange[i], self.vec_dim], dtype=tf.float32) for i in range(self.field_len)]
            Embedding_layer = tf.concat([tf.matmul(tf.slice(self.X,[0,self.field_arange[i]],[-1,self.field_arange[i+1]-self.field_arange[i]]), Embedding[i]) for i in range(self.field_len)], axis=1)
        x = Embedding_layer

        in_num = self.field_len * self.vec_dim

        for i in range(len(self.dnn_layers)):
            out_num = self.dnn_layers[i]
            w = tf.get_variable(name='w_%d'%i, shape=[in_num, out_num], dtype=tf.float32)
            b = tf.get_variable(name='b_%d'%i, shape=[out_num], dtype=tf.float32)
            x = tf.matmul(x, w) + b
            if out_num == 2:
                self.y_out = x + self.wide+b
            else:
                x = tf.nn.relu(x)
            in_num = out_num
    def add_loss(self):
        self.loss = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(labels=self.Y, logits=self.y_out))

    #计算accuracy
    def add_accuracy(self):
        # accuracy
        self.correct_prediction = tf.equal(tf.cast(tf.argmax(self.y_out,1), tf.float32), tf.cast(tf.argmax(self.Y,1), tf.float32))
        self.accuracy = tf.reduce_mean(tf.cast(self.correct_prediction, tf.float32))

        # self.auc_value = tf.metrics.auc(tf.argmax(self.y_out,1),tf.argmax(self.Y,1), curve='ROC')
    #训练
    def train(self):

        wide_opt = tf.train.FtrlOptimizer(learning_rate=self.linear_lr, l1_regularization_strength=self.l1_reg)
        wide_opt_train = wide_opt.minimize(loss=self.loss)

        deep_opt = tf.train.AdamOptimizer(learning_rate=self.dnn_lr)
        deep_opt_train = deep_opt.minimize(loss=self.loss)

        self.train_op = tf.group(wide_opt_train, deep_opt_train)

    #构建图
    def build_graph(self):
        self.add_input()
        self.inference()
        self.add_loss()
        self.add_accuracy()
        self.train()

def train_model(sess, model, X_train,Y_train,batch_size, epochs=100):
    num = len(X_train) // batch_size+1
    for step in range(epochs):
        print("epochs{0}:".format(step+1))
        for i in range(num):
            index = np.random.choice(len(X_train), batch_size)
            batch_x = X_train[index]
            batch_y = Y_train[index]
            feed_dict = {
    
    model.X: batch_x,
                         model.Y: batch_y}
            sess.run(model.train_op, feed_dict=feed_dict)


            # print("Iteration {0}: with minibatch  training loss = {1}"
            #       .format(step+1, loss))

            if (i+1)%100==0:
                loss ,accuracy,y_out= sess.run([model.loss,model.accuracy,model.y_out], feed_dict=feed_dict)
                auc = metrics.roc_auc_score(batch_y, y_out)
                print("Iteration {0}: with minibatch training loss = {1} accuracy = {2} auc={3}"
                      .format(step+1, loss,accuracy,auc))

def test_model(sess,model,X_test,Y_test):

    loss,y_out, accuracy= sess.run([model.loss, model.y_out,model.accuracy], feed_dict={
    
    model.X: X_test, model.Y: Y_test})

    print("loss={0} accuracy={1} auc={2}".format(loss,accuracy,metrics.roc_auc_score(Y_test, y_out)))

if __name__ == '__main__':
    X_train,Y_train,X_test,Y_test,field_arange,field_len=loadData()
    linear_lr = 0.001
    dnn_lr=0.001
    l1_reg=0.5
    batch_size = 128
    vec_dim = 10
    feature_length = X_train.shape[1]
    dnn_layers=[256,128,2]

    model = WideOrDeep(vec_dim, linear_lr, dnn_lr, l1_reg, feature_length, field_arange, field_len, dnn_layers)
    model.build_graph()


    with tf.Session() as sess:

        sess.run(tf.global_variables_initializer())
        print('start training...')
        train_model(sess,model,X_train,Y_train,batch_size,epochs=10)
        print('start testing...')
        test_model(sess,model,X_test,Y_test)