1. 自定义随时函数

下面的代码自定义了XGBoost的随时函数。XGBoost要求我们在自定义随时函数的时候，给出其一阶导函数和二阶导函数的值的表达式。同时，XGBoost也允许使用者给出自定义的评价函数，用以评价训练出来的模型的性能。

import xgboost as xgb
import numpy as np

# 1、xgBoost的基本使用
# 2、自定义损失函数的梯度和二阶导
# 3、binary:logistic/logitraw


# 定义f: theta * x
def log_reg(y_hat, y):
    p = 1.0 / (1.0 + np.exp(-y_hat))
    g = p - y.get_label()
    h = p * (1.0-p)
    return g, h


def error_rate(y_hat, y):
    return 'error', float(sum(y.get_label() != (y_hat > 0.5))) / len(y_hat)


if __name__ == "__main__":
    # 读取数据
    data_train = xgb.DMatrix('agaricus_train.txt')
    data_test = xgb.DMatrix('agaricus_test.txt')
    print data_train
    print type(data_train)

    # 设置参数
    param = {'max_depth': 3, 'eta': 1, 'silent': 1, 'objective': 'binary:logistic'} # logitraw
    # param = {'max_depth': 3, 'eta': 0.3, 'silent': 1, 'objective': 'reg:logistic'}
    watchlist = [(data_test, 'eval'), (data_train, 'train')]
    n_round = 7
    # bst = xgb.train(param, data_train, num_boost_round=n_round, evals=watchlist)
    bst = xgb.train(param, data_train, num_boost_round=n_round, evals=watchlist, obj=log_reg, feval=error_rate)

    # 计算错误率
    y_hat = bst.predict(data_test)
    y = data_test.get_label()
    print y_hat
    print y
    error = sum(y != (y_hat > 0.5))
    error_rate = float(error) / len(y_hat)
    print '样本总数：\t', len(y_hat)
    print '错误数目：\t%4d' % error
    print '错误率：\t%.5f%%' % (100*error_rate)

2. 读入数据的一些小操作

我们的原始数据在txt中是这样的

XGBoost其实可以直接读入这种数据，不需要进行下述操作。下述操作只是展示一下如何把这种数据改成我们常见的系数矩阵形式。

import xgboost as xgb
import numpy as np
import scipy.sparse
from sklearn.model_selection import train_test_split
from sklearn.linear_model import LogisticRegression
from sklearn.metrics import accuracy_score


def read_data(path):
    y = []
    row = []
    col = []
    values = []
    r = 0       # 首行
    for d in open(path):
        d = d.strip().split()      # 以空格分开
        y.append(int(d[0]))
        d = d[1:]
        for c in d:
            key, value = c.split(':')
            row.append(r)
            col.append(int(key))
            values.append(float(value))
        r += 1
    x = scipy.sparse.csr_matrix((values, (row, col))).toarray()
    y = np.array(y)
    return x, y


if __name__ == '__main__':
    x, y = read_data('agaricus_train.txt')
    x_train, x_test, y_train, y_test = train_test_split(x, y, random_state=1, train_size=0.6)

    # Logistic回归
    lr = LogisticRegression(penalty='l2')
    lr.fit(x_train, y_train.ravel())
    y_hat = lr.predict(x_test)
    print 'Logistic回归正确率：', accuracy_score(y_test, y_hat)

    # XGBoost
    data_train = xgb.DMatrix(x_train, label=y_train)
    data_test = xgb.DMatrix(x_test, label=y_test)
    watch_list = [(data_test, 'eval'), (data_train, 'train')]
    param = {'max_depth': 3, 'eta': 1, 'silent': 0, 'objective': 'multi:softmax', 'num_class': 3}
    bst = xgb.train(param, data_train, num_boost_round=4, evals=watch_list)
    y_hat = bst.predict(data_test)
    print 'XGBoost正确率：', accuracy_score(y_test, y_hat)

3. Titanic数据的实操

这段代码的主要关注点在于数据的预处理部分。展示了多种缺失值的处理办法，第一，直接用平均值或者是中位数进行缺失值替换。但是即使是这种比较简练的替换方法，也有可深入探索的地方，比如在替换年龄时，选取同一性别的人的年龄平均值进行替换，而不是用所有人的年龄平均值进行替换。第二，用机器学习模型，比如随机森林，进行缺失值替换。具体方法是，选取处缺失变量的其他变量作为解释变量，缺失变量作为反应变量，使用模型进行训练并做相应预测，进而替换缺失的部分。当然代码的数据预处理还要其他关注点，比如独热编码部分。

import xgboost as xgb
import numpy as np
from sklearn.linear_model import LogisticRegression
from sklearn.model_selection import train_test_split
from sklearn.ensemble import RandomForestRegressor
from sklearn.ensemble import RandomForestClassifier
from sklearn.metrics import accuracy_score
import pandas as pd
import csv


def show_accuracy(a, b, tip):
    acc = a.ravel() == b.ravel()
    acc_rate = 100 * float(acc.sum()) / a.size
    print '%s正确率：%.3f%%' % (tip, acc_rate)
    return acc_rate


def load_data(file_name, is_train):
    data = pd.read_csv(file_name)  # 数据文件路径
    # print 'data.describe() = \n', data.describe()

    # 性别
    data['Sex'] = data['Sex'].map({'female': 0, 'male': 1}).astype(int)

    # 补齐船票价格缺失值
    if len(data.Fare[data.Fare.isnull()]) > 0:
        fare = np.zeros(3)
        for f in range(0, 3):
            fare[f] = data[data.Pclass == f + 1]['Fare'].dropna().median()
        for f in range(0, 3):  # loop 0 to 2
            data.loc[(data.Fare.isnull()) & (data.Pclass == f + 1), 'Fare'] = fare[f]

    # 年龄：使用均值代替缺失值
    # mean_age = data['Age'].dropna().mean()
    # data.loc[(data.Age.isnull()), 'Age'] = mean_age
    if is_train:
        # 年龄：使用随机森林预测年龄缺失值
        print '随机森林预测缺失年龄：--start--'
        data_for_age = data[['Age', 'Survived', 'Fare', 'Parch', 'SibSp', 'Pclass']]
        age_exist = data_for_age.loc[(data.Age.notnull())]   # 年龄不缺失的数据
        age_null = data_for_age.loc[(data.Age.isnull())]
        # print age_exist
        x = age_exist.values[:, 1:]
        y = age_exist.values[:, 0]
        rfr = RandomForestRegressor(n_estimators=1000)
        rfr.fit(x, y)
        age_hat = rfr.predict(age_null.values[:, 1:])
        # print age_hat
        data.loc[(data.Age.isnull()), 'Age'] = age_hat
        print '随机森林预测缺失年龄：--over--'
    else:
        print '随机森林预测缺失年龄2：--start--'
        data_for_age = data[['Age', 'Fare', 'Parch', 'SibSp', 'Pclass']]
        age_exist = data_for_age.loc[(data.Age.notnull())]  # 年龄不缺失的数据
        age_null = data_for_age.loc[(data.Age.isnull())]
        # print age_exist
        x = age_exist.values[:, 1:]
        y = age_exist.values[:, 0]
        rfr = RandomForestRegressor(n_estimators=1000)
        rfr.fit(x, y)
        age_hat = rfr.predict(age_null.values[:, 1:])
        # print age_hat
        data.loc[(data.Age.isnull()), 'Age'] = age_hat
        print '随机森林预测缺失年龄2：--over--'

    # 起始城市
    data.loc[(data.Embarked.isnull()), 'Embarked'] = 'S'  # 保留缺失出发城市
    # data['Embarked'] = data['Embarked'].map({'S': 0, 'C': 1, 'Q': 2, 'U': 0}).astype(int)
    # print data['Embarked']
    embarked_data = pd.get_dummies(data.Embarked)
    print embarked_data
    # embarked_data = embarked_data.rename(columns={'S': 'Southampton', 'C': 'Cherbourg', 'Q': 'Queenstown', 'U': 'UnknownCity'})
    embarked_data = embarked_data.rename(columns=lambda x: 'Embarked_' + str(x))
    data = pd.concat([data, embarked_data], axis=1)
    print data.describe()
    data.to_csv('New_Data.csv')

    x = data[['Pclass', 'Sex', 'Age', 'SibSp', 'Parch', 'Fare', 'Embarked_C', 'Embarked_Q', 'Embarked_S']]
    # x = data[['Pclass', 'Sex', 'Age', 'SibSp', 'Parch', 'Fare', 'Embarked']]
    y = None
    if 'Survived' in data:
        y = data['Survived']

    x = np.array(x)
    y = np.array(y)

    # 思考：这样做，其实发生了什么？
    x = np.tile(x, (5, 1))
    y = np.tile(y, (5, ))
    if is_train:
        return x, y
    return x, data['PassengerId']


def write_result(c, c_type):
    file_name = 'Titanic.test.csv'
    x, passenger_id = load_data(file_name, False)

    if type == 3:
        x = xgb.DMatrix(x)
    y = c.predict(x)
    y[y > 0.5] = 1
    y[~(y > 0.5)] = 0

    predictions_file = open("Prediction_%d.csv" % c_type, "wb")
    open_file_object = csv.writer(predictions_file)
    open_file_object.writerow(["PassengerId", "Survived"])
    open_file_object.writerows(zip(passenger_id, y))
    predictions_file.close()


if __name__ == "__main__":
    x, y = load_data('Titanic.train.csv', True)
    x_train, x_test, y_train, y_test = train_test_split(x, y, test_size=0.25, random_state=1)
    #
    lr = LogisticRegression(penalty='l2')
    lr.fit(x_train, y_train)
    y_hat = lr.predict(x_test)
    lr_acc = accuracy_score(y_test, y_hat)
    # write_result(lr, 1)

    rfc = RandomForestClassifier(n_estimators=100)
    rfc.fit(x_train, y_train)
    y_hat = rfc.predict(x_test)
    rfc_acc = accuracy_score(y_test, y_hat)
    # write_result(rfc, 2)

    # XGBoost
    data_train = xgb.DMatrix(x_train, label=y_train)
    data_test = xgb.DMatrix(x_test, label=y_test)
    watch_list = [(data_test, 'eval'), (data_train, 'train')]
    param = {'max_depth': 6, 'eta': 0.8, 'silent': 1, 'objective': 'binary:logistic'}
             # 'subsample': 1, 'alpha': 0, 'lambda': 0, 'min_child_weight': 1}
    bst = xgb.train(param, data_train, num_boost_round=100, evals=watch_list)
    y_hat = bst.predict(data_test)
    # write_result(bst, 3)
    y_hat[y_hat > 0.5] = 1
    y_hat[~(y_hat > 0.5)] = 0
    xgb_acc = accuracy_score(y_test, y_hat)

    print 'Logistic回归：%.3f%%' % lr_acc
    print '随机森林：%.3f%%' % rfc_acc
    print 'XGBoost：%.3f%%' % xgb_acc