在Scikit Learn中拟合逻辑模型时，多项式特征无法正常工作

Kaggle有一个关于泰坦尼克号的介绍性数据科学问题，目标是预测乘客的生存机会，并给出一些关于他/她的信息（例如性别、年龄、客舱等级等）。我使用Scikit Learn在Python中实现了一个简单的逻辑回归模型，我正在探索添加"年龄"变量的更高阶因子。我按照Scikit Learn网站上的说明使用了PolynomialFeatures：

import pandas as pd
from sklearn import linear_model
import numpy as np
from sklearn import preprocessing
from sklearn.preprocessing import PolynomialFeatures
# Import titanic data
titanic = pd.read_csv("train.csv")
# Set the training set as 70% of the dataset and cross_validation set as remaining 30%
predictors = ["Age"]
training_set = titanic[predictors].iloc[range(0,int(titanic.shape[0]*0.7)),:]
cv_set = titanic[predictors].iloc[range(int(titanic.shape[0]*0.7),titanic.shape[0]),:]
training_actuals = titanic["Survived"].iloc[range(0,int(titanic.shape[0]*0.7))]
cv_actuals = titanic["Survived"].iloc[range(int(titanic.shape[0]*0.7),titanic.shape[0])]
# Create polynomial features
poly = PolynomialFeatures(degree=3)
training_set = poly.fit_transform(training_set)
cv_set = poly.fit_transform(cv_set)
# Fit a logistic regression model, predict values for training and cross-validation sets
alg = linear_model.LogisticRegression()
alg.fit(training_set, training_actuals)
cv_predictions = alg.predict(cv_set)
training_predictions = alg.predict(training_set)
# Measure and print accuracy of prediction over both training and cross-validation sets
cv_accuracy = len(cv_predictions[cv_predictions == np.array(cv_actuals)])/float(len(cv_predictions))
print "Prediction accuracy on cross-validation set is %s%%" % (cv_accuracy * 100)
training_accuracy = len(training_predictions[training_predictions == np.array(training_actuals)])/float(len(training_predictions))
print "Prediction accuracy on training set is %s%%" % (training_accuracy * 100)

当我添加年龄的平方特征（即多项式次数2）时，我在训练集上预测值的准确性提高了1-2个百分点，但当我进行次数3时，如上面的代码中所示，准确性实际上恢复到与线性情况相同（即次数=1）。从理论上讲，它应该略有改善，或者与度=2时保持不变。这种行为在所有更高的学位中也会继续。我是Scikit Learn的新手，如果能了解我做错了什么，我将不胜感激。