均值漂移算法(MeanShift)是一种旨在发现团(blobs)的聚类算法
核心思想
寻找核密度极值点并作为簇的质心,然后根据最近邻原则将样本点赋予质心
算法简介
核密度估计
根据样本分布估计在样本空间的每一点的密度。估计某点的密度时,核密度估计方法会考虑该点邻近区域的样本点的影响,邻近区域大小由带宽h决定,该参数对最终密度估计的影响非常大。通常采用高斯核: N ( x ) = 1 2 π h e x 2 2 h 2 N(x)=\frac{1}{\sqrt{2\pi}h}e^{\frac{x^2}{2h^2}} N(x)=2πh1e2h2x2
均值漂移
算法初始化一个质心(向量表示),每一步迭代都会朝着当前质心领域内密度极值方向漂移,方向就是密度上升最大的方向,即梯度方向(参考资料)。求导即得漂移向量的终点就是下一个质心: x k i + 1 = ∑ x j ∈ N K ( x j − x k i ) x j ∑ x j ∈ N K ( x j − x k i ) x_k^{i+1}=\frac{\sum_{x_j\in N}K(x_j-x_k^i)x_j}{\sum_{x_j\in N}K(x_j-x_k^i)} xki+1=∑xj∈NK(xj−xki)∑xj∈NK(xj−xki)xjN是当前质心的领域内样本集合。这个公式可以从另一方面理解:当前质心的邻域内样本点以核密度为权重的均值就是更新后的质心。
算法流程
- Input: 高斯核带宽,bin_seeding(是否对数据粗粒化), min_fre(可以作为起始质心的样本点领域内的最少样本数),阈值epsilon(漂移最小长度)
- Output: 样本簇标签
- Step1: 获取可以作为起始质心的样本点
- Step2: 对每个起始质心进行漂移,漂移终止条件就是漂移距离小于epsilon。若漂移结束最终的质心与已存在的质心距离小于带宽则合并
- Step3: 分类。将样本点归属到距离最近的质心中
代码
"""
meanshift聚类算法
核心思想:
寻找核密度极值点并作为簇的质心,然后根据最近邻原则将样本点赋予质心
"""
from collections import defaultdict
import numpy as np
import math
class MeanShift:
def __init__(self, band_width=2.0, min_fre=3, epsilon=None, bin_seeding=False, bin_size=None):
self.epsilon = epsilon if epsilon else 1e-3 * band_width
self.bin_size = bin_size if bin_size else self.band_width
self.band_width = band_width
self.min_fre = min_fre # 可以作为起始质心的球体内最少的样本数目
self.bin_seeding = bin_seeding
self.radius2 = self.band_width ** 2 # 高维球体半径的平方
self.N = None
self.labels = None
self.centers = []
def init_param(self, data):
# 初始化参数
self.N = data.shape[0]
self.labels = -1 * np.ones(self.N)
return
def get_seeds(self, data):
# 获取可以作为起始质心的点(seed)
if not self.bin_seeding:
return data
seed_list = []
seeds_fre = defaultdict(int)
for sample in data:
seed = tuple(np.round(sample / self.bin_size)) # 将数据粗粒化,以防止非常近的样本点都作为起始质心
seeds_fre[seed] += 1
for seed, fre in seeds_fre.items():
if fre >= self.min_fre:
seed_list.append(np.array(seed))
if not seed_list:
raise ValueError('the bin size and min_fre are not proper')
if len(seed_list) == data.shape[0]:
return data
return np.array(seed_list) * self.bin_size
def euclidean_dis2(self, center, sample):
# 计算均值点到每个样本点的欧式距离(平方)
delta = center - sample
return delta @ delta
def gaussian_kel(self, dis2):
# 计算高斯核
return 1.0 / self.band_width * (2 * math.pi) ** (-1.0 / 2) * math.exp(-dis2 / (2 * self.band_width ** 2))
def shift_center(self, current_center, data):
# 计算下一个漂移的坐标
denominator = 0 # 分母
numerator = np.zeros_like(current_center) # 分子, 一维数组形式
for sample in data:
dis2 = self.euclidean_dis2(current_center, sample)
if dis2 <= self.radius2:
d = self.gaussian_kel(dis2)
denominator += d
numerator += d * sample
if denominator > 0:
return numerator / denominator
else:
return None
def classify(self, data):
# 根据最近邻将数据分类到最近的簇中
center_arr = np.array(self.centers)
for i in range(self.N):
delta = center_arr - data[i]
dis2 = np.sum(delta * delta, axis=1)
self.labels[i] = np.argmin(dis2)
return
def fit(self, data):
# 训练主函数
self.init_param(data)
seed_list = self.get_seeds(data)
for seed in seed_list:
bad_seed = False
current_center = seed
# 进行一次独立的均值漂移
while True:
next_center = self.shift_center(current_center, data)
if next_center is None:
bad_seed = True
break
delta_dis = np.linalg.norm(next_center - current_center, 2)
if delta_dis < self.epsilon:
break
current_center = next_center
if not bad_seed:
# 若该次漂移结束后,最终的质心与已存在的质心距离小于带宽,则合并
for i in range(len(self.centers)):
if np.linalg.norm(current_center - self.centers[i], 2) < self.band_width:
break
else:
self.centers.append(current_center)
self.classify(data)
return
if __name__ == '__main__':
from sklearn.datasets import make_blobs
data, label = make_blobs(n_samples=500, centers=5, cluster_std=1.2, random_state=7)
MS = MeanShift(band_width=3, min_fre=3, bin_size=4, bin_seeding=True)
MS.fit(data)
labels = MS.labels
print(MS.centers, np.unique(labels))
import matplotlib.pyplot as plt
from itertools import cycle
def visualize(data, labels):
color = 'bgrymk'
unique_label = np.unique(labels)
for col, label in zip(cycle(color), unique_label):
partial_data = data[np.where(labels == label)]
plt.scatter(partial_data[:, 0], partial_data[:, 1], color=col)
plt.show()
return
visualize(data, labels)
我的GitHub
注:如有不当之处,请指正。