# week12

**Repository Path**: yaotengdong/week12

## Basic Information

- **Project Name**: week12
- **Description**: No description available
- **Primary Language**: Unknown
- **License**: Not specified
- **Default Branch**: master
- **Homepage**: None
- **GVP Project**: No

## Statistics

- **Stars**: 0
- **Forks**: 0
- **Created**: 2019-07-21
- **Last Updated**: 2020-12-19

## Categories & Tags

**Categories**: Uncategorized

**Tags**: None

## README

## 1、采用Wide and Deep模型，对Criteo提供的Kaggle竞赛数据进行CTR预估。 

## 2、数据描述： 
数据共包含11天的数据，其中10天为训练数据train，1天为测试数据test。<br> 
文件说明 <br>
train.csv: 训练数据。 <br>
eval.csv：测试数据 <br>
字段说明：字段已进行脱敏处理 <br>
I1-I13: 整数型特征 <br>
C1-C26：类别型特征，已进行Hash编码 <br>
clicked：是否被点击 <br>

## 3、评价标准： <br>
项目采用logloss评价模型性能。令y_i为第i个样本真正的标签，y ̂_i为模型对第i个的预测值，N为样本数据，则logloss为 <br>
logloss=-1/N ∑_(i=1)^N▒〖y_i log⁡(y ̂_i)〖(1-y〗_i)log⁡(1-y ̂_i)〗 。 <br>

其他参考资料： <br>
TensorFlow Wide And Deep 模型详解与应用：https://cloud.tencent.com/developer/article/1143316 <br>
中文版：https://www.helplib.com/GitHub/article_153816 <br>
Wide and Deep模型分类例子：https://github.com/tensorflow/models/tree/master/official/wide_deep <br>
https://github.com/gutouyu/ML_CIA/tree/master/Wide&Deep <br>
一个包含多个深度模型的用于CTR的工具包：DeepCTR <br>
https://github.com/shenweichen/DeepCTR <br>
中文版：https://zhuanlan.zhihu.com/p/53231955<br>

## 代码实现：

### 1.查看原始数据，没有label, 所以根据说明给数据进行分类处理，数据数据为I1 - I13，类别型数据为C1 - C26
```
CONTINUOUS_COLUMNS =  ["I"+str(i) for i in range(1,14)] # 1-13 inclusive
CATEGORICAL_COLUMNS = ["C"+str(i) for i in range(1,27)] # 1-26 inclusive
LABEL_COLUMN = ["clicked"]

TRAIN_DATA_COLUMNS = LABEL_COLUMN + CONTINUOUS_COLUMNS + CATEGORICAL_COLUMNS
# TEST_DATA_COLUMNS = CONTINUOUS_COLUMNS + CATEGORICAL_COLUMNS

FEATURE_COLUMNS = CONTINUOUS_COLUMNS + CATEGORICAL_COLUMNS
```
### 2. 创建特征列

#### 创建类别型特征列


```
wide_columns = []
for name in CATEGORICAL_COLUMNS:
    wide_columns.append(tf.contrib.layers.sparse_column_with_hash_bucket(
            name, hash_bucket_size=1000))
```


#### 创建数值型特征列


```
deep_columns = []
for name in CONTINUOUS_COLUMNS:
    deep_columns.append(tf.contrib.layers.real_valued_column(name))
```

#### 创建数值与类别组合的特征列（把类别型特征列嵌入到数值型特征）


```
for col in wide_columns:
    deep_columns.append(tf.contrib.layers.embedding_column(col, 
                                                           dimension=8))
```

### 3. 创建训练模型

创建三类模型：1）类别型数据分类的线性分类器；2）数值型特征使用的深度神经网络分类器；3）组合数据使用的组合线性和深度分类器<br>


```
 # Linear Classifier
    if model_type == 'WIDE':
        m = tf.contrib.learn.LinearClassifier(
            model_dir=model_dir, 
            feature_columns=wide_columns)

    # Deep Neural Net Classifier
    if model_type == 'DEEP':
        m = tf.contrib.learn.DNNClassifier(
            model_dir=model_dir,
            feature_columns=deep_columns,
            hidden_units=[100, 50, 25])

    # Combined Linear and Deep Classifier
    if model_type == 'WIDE_AND_DEEP':
        m = tf.contrib.learn.DNNLinearCombinedClassifier(
            model_dir=model_dir,
            linear_feature_columns=wide_columns,
            dnn_feature_columns=deep_columns,
            dnn_hidden_units=[100, 70, 50, 25],
            config=runconfig)
        
```

### 4. 利用生成好的模型，对数据进行训练：


```
%%time

# This can be found with
# wc -l train.csv
train_sample_size = 800000
train_steps = train_sample_size/BATCH_SIZE # 8000/40 = 200

model_w_n_d.fit(input_fn=generate_input_fn(train_file, BATCH_SIZE), steps=train_steps)
model_w.fit(input_fn=generate_input_fn(train_file, BATCH_SIZE), steps=train_steps)
model_d.fit(input_fn=generate_input_fn(train_file, BATCH_SIZE), steps=train_steps)

print('fit done')
```

### 5. 对eval数据进行评估验证：

1）Wide and Deep 模型

```
%%time

eval_sample_size = 200000 # this can be found with a 'wc -l eval.csv'
eval_steps = eval_sample_size/BATCH_SIZE # 2000/40 = 50

results = model_w_n_d.evaluate(input_fn=generate_input_fn(eval_file), 
                     steps=eval_steps)
print('evaluate done')

print('Accuracy: %s' % results['accuracy'])
print(results)
```
结果：'loss': 0.4962089, 'accuracy': 0.768985<br>

2）Wide 模型


```
results = model_w.evaluate(input_fn=generate_input_fn(eval_file), 
                     steps=eval_steps)
print('evaluate done')

print('Accuracy: %s' % results['accuracy'])
print(results)
```
结果：'loss': 0.50244325, 'accuracy': 0.766125<br>


3）Deep 模型


```
results = model_d.evaluate(input_fn=generate_input_fn(eval_file), 
                     steps=eval_steps)
print('evaluate done')

print('Accuracy: %s' % results['accuracy'])
print(results)
```
结果：'loss': 0.5411806, 'accuracy': 0.75249<br>

所以综合验证，Wide and Deep 模型的效果（loss:0.4962089）比单一的模型效果好很多(wide loss:0.50244325, deep loss: 0.5411806)。