朴素贝叶斯分类原理
对于给定的训练数据集,首先基于特征条件独立假设学习输入/输出的联合概率分布;然后基于此模型,对给定的输入(x),利用贝叶斯定理求出后验概率最大的输出(y)。
特征独立性假设:在利用贝叶斯定理进行预测时,我们需要求解条件概率(P(x|y_k)=P(x_1,x_2,...,x_n|y_k)P(x|y_k)=P(x_1,x_2,...,x_n|y_k)),它的参数规模是指数数量级别的,假设第i维特征可取值的个数有(T_i)个,类别取值个数为k个,那么参数个数为:(kprod_{i=1}^nT_i)。这显然不可行,所以朴素贝叶斯算法对条件概率分布作出了独立性的假设,实际上是为了简化计算。
import numpy as np
import math
from sklearn.datasets import load_iris
from sklearn.model_selection import train_test_split
from collections import Counter
从sklearn数据集中加载鸢尾花分类数据集
iris = load_iris()
X, Y = iris.data, iris.target
X_train, X_test, Y_train, Y_test = train_test_split(X, Y, test_size=0.3)
print('X_train[0]: {}'.format(X_train[0]))
print('Y_train[0]: {}'.format(Y_train[0]))
# 查看训练集各个类别的数量
for l in set(Y_train):
print('label: %s ,count: %d' % (l, len(Y_train[Y_train==l])))
代码输出:
X_train[0]: [5.2 3.5 1.5 0.2]
Y_train[0]: 0
label: 0 ,count: 35
label: 1 ,count: 32
label: 2 ,count: 38
高斯模型的朴素贝叶斯:
对于取值是连续型的特征变量,用离散型特征的求解方法时会有很多特征取值的条件概率为0,所以我们使用高斯模型的朴素贝叶斯,它假设每一维特征都服从高斯分布。即:
[P(x_i | y_k)=frac{1}{sqrt{2pi}sigma_{y_k,i}}exp(-frac{(x_i-mu_{y_k,i})^2}{2sigma^2_{y_k,i}})
]
(mu_{y_k,i})是分类为(y_k)的样本中,第(i)维特征取值的均值;(sigma_{y_k,i}^2)为其方差
class GaussianNaiveBayes:
def __init__(self):
self.parameters = {}
self.prior = {}
# 训练过程就是求解先验概率和高斯分布参数的过程
# X:(样本数,特征维度) Y:(样本数,)
def fit(self, X, Y):
self._get_prior(Y) # 计算先验概率
labels = set(Y)
for label in labels:
samples = X[Y==label]
# 计算高斯分布的参数:均值和标准差
means = np.mean(samples, axis=0)
stds = np.std(samples, axis=0)
self.parameters[label] = {
'means': means,
'stds': stds
}
# x:单个样本
def predict(self, x):
probs = sorted(self._cal_likelihoods(x).items(), key=lambda x:x[-1]) # 按概率从小到大排序
return probs[-1][0]
# 计算模型在测试集的准确率
# X_test:(测试集样本个数,特征维度)
def evaluate(self, X_test, Y_test):
true_pred = 0
for i, x in enumerate(X_test):
label = self.predict(x)
if label == Y_test[i]:
true_pred += 1
return true_pred / len(X_test)
# 计算每个类别的先验概率
def _get_prior(self, Y):
cnt = Counter(Y)
for label, count in cnt.items():
self.prior[label] = count / len(Y)
# 高斯分布
def _gaussian(self, x, mean, std):
exponent = math.exp(-(math.pow(x - mean, 2)/(2 * math.pow(std, 2))))
return (1 / (math.sqrt(2 * math.pi) * std)) * exponent
# 计算样本x属于每个类别的似然概率
def _cal_likelihoods(self, x):
likelihoods = {}
for label, params in self.parameters.items():
means = params['means']
stds = params['stds']
prob = self.prior[label]
# 计算每个特征的条件概率,P(xi|yk)
for i in range(len(means)):
prob *= self._gaussian(x[i], means[i], stds[i])
likelihoods[label] = prob
return likelihoods
在测试集上评估分类器:
gussian_nb = GaussianNaiveBayes()
gussian_nb.fit(X_train, Y_train)
print('样本[4.4, 3.2, 1.3, 0.2]的预测结果: %d' % gussian_nb.predict([4.4, 3.2, 1.3, 0.2]))
print('测试集的准确率: %f' % gussian_nb.evaluate(X_test, Y_test))
代码输出:
样本[4.4, 3.2, 1.3, 0.2]的预测结果: 0
测试集的准确率: 0.955556
与scikit-learn的实现对比
from sklearn.naive_bayes import GaussianNB
clf = GaussianNB()
clf.fit(X_train, Y_train)
print('(sklearn)样本[4.4, 3.2, 1.3, 0.2]的预测结果: %d' % clf.predict([[4.4, 3.2, 1.3, 0.2]])[0])
print('(sklearn)测试集的准确率: %f' % clf.score(X_test, Y_test))
代码输出:
(sklearn)样本[4.4, 3.2, 1.3, 0.2]的预测结果: 0
(sklearn)测试集的准确率: 0.955556