Kaggle题目 nu-cs6220-assignment-1
题目地址如下：
https://www.kaggle.com/c/nu-cs6220-assignment-1/overview
这是个二分类任务，需要预测一个人的收入，分为两类：收入大于50K，或是小于50K。
1. 查看数据结构
下载数据后，先大致了解数据：
raw_data = load_data('nu-cs/training.txt')
raw_data.head()
可以看到没有header，根据题目对数据的说明，给它们分配header：
header = ['age', 'workclass', 'fnlwgt', 'education', 'education-num', 'marital-status', 'occupation', 'relationship', 'race', 'sex', 'capital-gain', 'capital-loss', 'hours-per-week', 'native-country', 'salary'] raw_data.columns = header raw_data.head()
在这个问题中，label为’salary’，这里它是一个离散变量，可以看到其中一个值是 ‘ <=50K’。进一步查看一下这个label包含的离散值：
raw_data['salary'].value_counts()
可以看到仅包含两类，且无缺失值或异常值。
继续查看数据集描述：
raw_data.info()
一共15个特征，6个为连续型，9个为离散型。数据条目为32560，每个特征均包含32560，但这个并不能说明数据集中没有缺失值，根据题目描述，缺失值已由 ? 代替。
2. 数值型特征
对于数值型特征，看一下统计数据：
raw_data.describe()
以及直方图：
结合这两组信息，我们可以看到有几点需要注意的地方：
- Capital-gain与capital-loss 大部分的值都是0，但是最大值却非常大，导致方差较大。
- 有些直方图是长尾分布，可以尝试将它们的分布转为钟型分布
- 部分特征需要进行分箱处理
下面先依次处理连续型变量。
2.1. Age特征
首先对于age特征，对它进行分箱并查看它们的相关性：
raw_data['age_band'] = pd.cut(raw_data['age'], 5)
raw_data[['age_band', 'salary']].groupby(['age_band'], as_index=False).mean().sort_values(by='age_band', ascending=True)
然后根据此分段，使用有序值替换 age 的值：
raw_data.loc[raw_data['age'] <= 31.6, 'age'] = 0 raw_data.loc[(raw_data['age'] <= 46.2) & (raw_data['age'] > 31.6), 'age'] = 1 raw_data.loc[(raw_data['age'] <= 60.8) & (raw_data['age'] > 46.2), 'age'] = 2 raw_data.loc[(raw_data['age'] <= 75.4) & (raw_data['age'] > 60.8), 'age'] = 3 raw_data.loc[(raw_data['age'] <= 90.0) & (raw_data['age'] > 75.4), 'age'] = 4
检查结果：
raw_data['age'].value_counts() 1 12210 0 11460 2 6558 3 2091 4 241 Name: age, dtype: int64
最后丢弃age_band 特征：
raw_data = raw_data.drop(['age_band'], axis=1)
raw_data.head()
2.2. fnlwgt特征
这个特征的问题在于：数值范围和方差都非常的大。
首先看它的直方图：
raw_data['fnlwgt'].hist(bins=100)
可以看到取值范围非常广，且类长尾分布。我们对它取对数，然后再观察它的直方图：
import numpy as np
raw_data['log_fnlwgt'] = raw_data['fnlwgt'].apply(np.log)
raw_data[['log_fnlwgt','fnlwgt']].hist(bins=100)
可以看到取对数后更接近钟型分布。最后，丢弃 log_fnlwgt，并直接在fnlwgt 上做变换：
raw_data.drop(['log_fnlwgt'], axis=1)
raw_data['fnlwgt'] = raw_data['fnlwgt'].apply(np.log)
2.3. Education-num
对于 education-num，它的取值虽然是数值型，但训练集中为有限集：
raw_data['education-num'].value_counts() 9 10501 10 7291 13 5354 14 1723 11 1382 7 1175 12 1067 6 933 4 646 15 576 5 514 8 433 16 413 3 333 2 168 1 51 Name: education-num, dtype: int64
同样对它使用区间量化，同 age 特征。过程在此不赘述，处理后的结果：
raw_data['education-num'].value_counts() 2 18225 3 7803 4 2712 1 2622 0 1198 Name: education-num, dtype: int64
2.4. capital-gain 与 capital-loss
raw_data[['capital-gain', 'capital-loss']].hist()
这两个特征的特点是：大部分值都为0，少部分值特别大。对这两个特征，采用二值化处理：
raw_data[['capital-gain', 'capital-loss']] = (raw_data[['capital-gain', 'capital-loss']] > 0) * 1
处理后结果为：
raw_data['capital-gain'].value_counts() 0 29849 1 2711 Name: capital-gain, dtype: int64 raw_data['capital-loss'].value_counts() 0 31041 1 1519 Name: capital-loss, dtype: int64
2.5. hours-per-week
此特征的图像类似为钟型分布，可以直接做标准化处理，或是做分桶处理均可，在此做了分桶处理，过程不追溯。
3. 离散特征
对于离散型特征，我们会用One-Hot 编码处理。首先我们清理缺失值：
workclass中存在 1836 条缺失值：
raw_data['workclass'].value_counts() Private 22696 Self-emp-not-inc 2541 Local-gov 2093 ? 1836 State-gov 1297 Self-emp-inc 1116 Federal-gov 960 Without-pay 14 Never-worked 7 Name: workclass, dtype: int64
occupation 中存在 1843 条缺失值：
raw_data['occupation'].value_counts() Prof-specialty 4140 Craft-repair 4099 Exec-managerial 4066 Adm-clerical 3769 Sales 3650 Other-service 3295 Machine-op-inspct 2002 ? 1843 Transport-moving 1597 Handlers-cleaners 1370 Farming-fishing 994 Tech-support 928 Protective-serv 649 Priv-house-serv 149 Armed-Forces 9 Name: occupation, dtype: int64
native-country 中存在583 条缺失值：
raw_data['native-country'].value_counts() United-States 29169 Mexico 643 ? 583 Philippines 198 …
对于这些缺失值，我们简单地使用众数来填充这个缺失值：
freq_workclass = raw_data.workclass.mode()[0] raw_data.loc[(raw_data['workclass'] == ' ?'), 'workclass'] = freq_workclass freq_occupation = raw_data.occupation.mode()[0] raw_data.loc[(raw_data['occupation'] == ' ?'), 'occupation'] = freq_workclass freq_nativecountry = raw_data['native-country'].mode()[0] raw_data.loc[(raw_data['native-country'] == ' ?'), 'native-country'] = freq_nativecountry
补全缺失值后，我们可以对它们应用one-hot 编码。不过对于native-country 特征，里面包含的离散值类别过多，若是使用 one-hot 编码，则势必会造成特征维度大大增加。这里我们用更少的特征去对它们进行替换：
raw_data.loc[raw_data['native-country'] == ' Scotland', 'native-country'] = 'UK' raw_data.loc[raw_data['native-country'] == ' United-States', 'native-country'] = 'US' raw_data.loc[raw_data['native-country'] == ' Mexico', 'native-country'] = 'South-America' raw_data.loc[raw_data['native-country'] == ' Jamaica', 'native-country'] = 'South-America' raw_data.loc[raw_data['native-country'] == ' Philippines', 'native-country'] = 'Asia' raw_data.loc[raw_data['native-country'] == ' Germany', 'native-country'] = 'Euro' raw_data.loc[raw_data['native-country'] == ' Canada', 'native-country'] = 'North-America' raw_data.loc[raw_data['native-country'] == ' Puerto-Rico', 'native-country'] = 'South-America' raw_data.loc[raw_data['native-country'] == ' El-Salvador', 'native-country'] = 'South-America' raw_data.loc[raw_data['native-country'] == ' India', 'native-country'] = 'Asia' raw_data.loc[raw_data['native-country'] == ' Cuba', 'native-country'] = 'South-America' raw_data.loc[raw_data['native-country'] == ' England', 'native-country'] = 'UK' raw_data.loc[raw_data['native-country'] == ' Italy', 'native-country'] = 'Euro' raw_data.loc[raw_data['native-country'] == ' Dominican-Republic', 'native-country'] = 'South-America' raw_data.loc[raw_data['native-country'] == ' Vietnam', 'native-country'] = 'Asia' raw_data.loc[raw_data['native-country'] == ' Guatemala', 'native-country'] = 'South-America' raw_data.loc[raw_data['native-country'] == ' Poland', 'native-country'] = 'Euro' raw_data.loc[raw_data['native-country'] == ' Columbia', 'native-country'] = 'South-America' raw_data.loc[raw_data['native-country'] == ' Haiti', 'native-country'] = 'South-America' raw_data.loc[raw_data['native-country'] == ' Portugal', 'native-country'] = 'Euro' raw_data.loc[raw_data['native-country'] == ' Greece', 'native-country'] = 'Euro' raw_data.loc[raw_data['native-country'] == ' France', 'native-country'] = 'Euro' raw_data.loc[raw_data['native-country'] == ' Ireland', 'native-country'] = 'Euro' raw_data.loc[raw_data['native-country'] == ' Holand-Netherlands', 'native-country'] = 'Euro' raw_data.loc[raw_data['native-country'] == ' China', 'native-country'] = 'Asia' raw_data.loc[raw_data['native-country'] == ' Japan', 'native-country'] = 'Asia' raw_data.loc[raw_data['native-country'] == ' Taiwan', 'native-country'] = 'Asia' raw_data.loc[raw_data['native-country'] == ' Hong', 'native-country'] = 'Asia' raw_data.loc[raw_data['native-country'] == ' Nicaragua', 'native-country'] = 'South-America' raw_data.loc[raw_data['native-country'] == ' Peru', 'native-country'] = 'South-America' raw_data.loc[raw_data['native-country'] == ' Ecuador', 'native-country'] = 'South-America' raw_data.loc[raw_data['native-country'] == ' Cambodia', 'native-country'] = 'Asia' raw_data.loc[raw_data['native-country'] == ' Thailand', 'native-country'] = 'Asia' raw_data.loc[raw_data['native-country'] == ' Laos', 'native-country'] = 'Asia' raw_data.loc[raw_data['native-country'] == ' Trinadad&Tobago', 'native-country'] = 'South-America' raw_data.loc[raw_data['native-country'] == ' Yugoslavia', 'native-country'] = 'Euro' raw_data.loc[raw_data['native-country'] == ' Honduras', 'native-country'] = 'South-America' raw_data.loc[raw_data['native-country'] == ' Hungary', 'native-country'] = 'Euro' raw_data.loc[raw_data['native-country'] == ' Iran', 'native-country'] = 'Middle-East' raw_data.loc[raw_data['native-country'] == ' South', 'native-country'] = 'South-America' raw_data['native-country'].value_counts() US 29752 South-America 1481 Asia 628 Euro 419 North-America 121 UK 102 Middle-East 43 Outlying-US(Guam-USVI-etc) 14 Name: native-country, dtype: int64
另一个可以进一步处理的特征是workclass，可以看到 workclass里的类别为：
Raw_data['workclass'].value_counts() Private 24532 Self-emp-not-inc 2541 Local-gov 2093 State-gov 1297 Self-emp-inc 1116 Federal-gov 960 Without-pay 14 Never-worked 7 Name: workclass, dtype: int64
其中 Without-pay 与 Nerver-worked 数量都比较少，也意思接近，我们将它作为一个类别处理：
def change_workclass(df): df.loc[df['workclass'] == ' Without-pay', 'workclass'] = 'No-pay' df.loc[df['workclass'] == ' Never-worked', 'workclass'] = 'No-pay'
4. 数据中心化、标准化以及One-Hot编码
在连续性变量与离散型变量均处理完毕后，将特征数据与label数据分离：
def get_data_label(df, label): dataset = df.drop(label, axis=1) labels = df[label].copy() return dataset, labels dataset, labels = get_data_label(raw_data, 'salary')
然后分别对数值型做中心化与标准化，离散值做one-hot编码：
from sklearn.pipeline import Pipeline from sklearn.preprocessing import StandardScaler from sklearn.impute import SimpleImputer from sklearn.pipeline import Pipeline from sklearn.preprocessing import StandardScaler from sklearn.impute import SimpleImputer from sklearn.compose import ColumnTransformer num_pipeline = Pipeline([ ('imputer', SimpleImputer(strategy='median')), ('std_scaler', StandardScaler()), ]) full_pipeline = ColumnTransformer([ ('num', num_pipeline, num_attributes), ('cat', OneHotEncoder(), cat_attributes), ]) nu_cs_prepared = full_pipeline.fit_transform(dataset)
模型训练
首先我们用sk-learn提供的几个模型训练：
from sklearn.model_selection import cross_val_score cross_val_score(tree_clf, nu_cs_prepared, labels, cv=10, scoring='accuracy') array([0.77955173, 0.78194103, 0.78961916, 0.77610565, 0.79299754, 0.7779484 , 0.78808354, 0.79207617, 0.79391892, 0.77695853]) from sklearn.svm import LinearSVC svm_clf = LinearSVC(C=3, loss="hinge") cross_val_score(svm_clf, nu_cs_prepared, labels, cv=10, scoring='accuracy') array([0.83911575, 0.83814496, 0.84520885, 0.82800983, 0.84029484, 0.84459459, 0.83630221, 0.84735872, 0.84490172, 0.83870968]) # logistic regression logreg = LogisticRegression() cross_val_score(logreg, nu_cs_prepared, labels, cv=10, scoring='accuracy') array([0.84280012, 0.84029484, 0.84981572, 0.83169533, 0.84398034, 0.84797297, 0.84029484, 0.8470516 , 0.85104423, 0.84423963]) from sklearn.ensemble import RandomForestClassifier rnd_clf = RandomForestClassifier(n_estimators=500, max_leaf_nodes=16, n_jobs=-1) cross_val_score(rnd_clf, nu_cs_prepared, labels, cv=10, scoring='accuracy') array([0.82222905, 0.82493857, 0.83015971, 0.82340295, 0.82831695, 0.82985258, 0.82186732, 0.83476658, 0.8252457 , 0.82519201])
可以看到表现最好的是SVM和LR。下面选择SVM，进行超参数搜索：
from sklearn.model_selection import GridSearchCV param_grid = [ {'C':[1, 3, 10, 30], 'loss':['hinge'], 'dual':[True]} ] svm_clf = LinearSVC() grid_search = GridSearchCV(svm_clf, param_grid, cv=5, scoring='accuracy', return_train_score=True) grid_search.fit(nu_cs_prepared, labels) grid_search.best_estimator_ LinearSVC(C=30, class_weight=None, dual=True, fit_intercept=True, intercept_scaling=1, loss='hinge', max_iter=1000, multi_class='ovr', penalty='l2', random_state=None, tol=0.0001, verbose=0) grid_search.best_score_ 0.832063882063882
可以看到最好的模型C=30（所以还可以往上调整C进一步搜索），此时的准确率为83%，但仍比不上 LR 的准确率。
再试试对随机森林的超参数搜索：
from sklearn.model_selection import GridSearchCV param_grid = [ {'n_estimators':[3, 10, 30], 'max_features':[2, 4, 6, 8]}, {'bootstrap':[False], 'n_estimators':[3, 10], 'max_features':[2, 3, 4]}, ] forest_clf = RandomForestClassifier() grid_search = GridSearchCV(forest_clf, param_grid, cv=5, scoring='accuracy', return_train_score=True) grid_search.fit(nu_cs_prepared, labels)
表现最好的参数为：
grid_search.best_params_ {'max_features': 8, 'n_estimators': 30}
最高分为：
0.8196253071253071
效果仍逊色于LR 的平均0.84 左右，下一章再试试 sagemaker 对模型进行训练。