简介

应用

概念

聚类算法与分类算法最大的区别

聚类算法是⽆监督的学习算法，⽽分类算法属于监督的学习算法。

API

案例

流程分析

导入依赖

import matplotlib.pyplot as plt
from sklearn.datasets import make_blobs
from sklearn.cluster import KMeans
from sklearn.metrics import calinski_harabasz_score

创建数据

# X为样本特征，Y为样本簇类别， 共1000个样本，每个样本2个特征，共4个簇， 
# 簇中⼼在[-1,-1], [0,0],[1,1], [2,2]， 簇⽅差分别为[0.4, 0.2, 0.2, 0.2]
x,y = make_blobs(n_samples=1000,
                 n_features=2,
                 centers=[[-1,-1],[0,0],[1,1],[2,2]],
                 cluster_std=[0.4,0.2,0.2,0.2],
                 random_state=9)
x,y

数据可视化

plt.scatter(x[:,0],x[:,1],marker="o")
plt.show()

kmeans 训练，可视化,聚类为2

y_pre = KMeans(n_clusters=2,random_state=9).fit_predict(x)
# 可视化
plt.scatter(x[:,0],x[:,1],marker="o",c=y_pre)
plt.show()
# 用ch_scole 查看最后效果
print(calinski_harabasz_score(x,y_pre))

kmeans 训练，可视化,聚类为3

y_pre = KMeans(n_clusters=3,random_state=9).fit_predict(x)
# 可视化
plt.scatter(x[:,0],x[:,1],marker="o",c=y_pre)
plt.show()
# 用ch_scole 查看最后效果
print(calinski_harabasz_score(x,y_pre))

kmeans 训练，可视化,聚类为4

y_pre = KMeans(n_clusters=4,random_state=9).fit_predict(x)
# 可视化
plt.scatter(x[:,0],x[:,1],marker="o",c=y_pre)
plt.show()
# 用ch_scole 查看最后效果
print(calinski_harabasz_score(x,y_pre))

聚类算法实现流程

K-means实现流程

模型评估

误差平⽅和(SSE\The sum of squares due to error)

“肘”⽅法(Elbowmethod)—K值确定

轮廓系数法（SilhouetteCoefficient）

CH系数（Calinski-HarabaszIndex）

小结

算法优化

K-means优点

K-means缺点

Canopy算法配合初始聚类

Canopy算法配合初始聚类实现流程

优缺点

K-means++

二分k-means

实现流程

k-medoids(K-中心聚类算法)

Kernel k-means (了解)

ISODATA（了解）

Mini Batch K-Means（了解）

特征降维

降维-特征选择

方法

低方差特征过滤

API

threshold：阀值方差

import pandas as pd
from sklearn.feature_selection import VarianceThreshold

data = pd.read_csv("factor_returns.csv")
print(data.head())
print(data.shape)
# 实例化对象
transfer = VarianceThreshold(threshold=2)
# 转换
transfer_data = transfer.fit_transform(data.iloc[:,1:10])
print(transfer_data)
print(data.iloc[:,1:10].shape)
print(transfer_data.shape)

皮尔逊相关系数（Pearson Correlation Coefficient）

例子

特点

API

from scipy.stats import pearsonr
x1=[12.5,15.3,23.2,26.4,33.5,34.4,39.4,45.2,55.4,60.9]
x2=[21.2,23.9,32.9,34.1,42.5,43.2,49.0,52.8,59.4,63.5]
ret = pearsonr(x1,x2)
print("皮尔逊相关系数的结果是:\n",ret)

斯⽪尔曼相关系数(RankIC)

例子

特点

API

案例

from scipy.stats import spearmanr
x1=[12.5,15.3,23.2,26.4,33.5,34.4,39.4,45.2,55.4,60.9]
x2=[21.2,23.9,32.9,34.1,42.5,43.2,49.0,52.8,59.4,63.5]
ret = spearmanr(x1,x2)
print("斯⽪尔曼相关系数的结果是:\n",ret)

降维-主成分分析（可以理解为特征提取）

API

from sklearn.decomposition import PCA
data=[[2,8,4,5],[6,3,0,8],[5,4,9,1]]
# pca小数保留百分比
transfer = PCA(n_components=0.9)
trans_data = transfer.fit_transform(data)
print("保留0.9的数据最后维度:\n",trans_data)

# pca保留3列
transfer = PCA(n_components=3)
trans_data = transfer.fit_transform(data)
print("保留3列的数据最后维度:\n",trans_data)

pca,K-means 实现用户对物品类别的喜好划分案例

import pandas as pd
from sklearn.decomposition import PCA
from sklearn.cluster import KMeans
from sklearn.metrics import silhouette_score

获取数据

order_product = pd.read_csv("order_products__prior.csv")
products = pd.read_csv("products.csv")
orders = pd.read_csv("orders.csv")
aisles = pd.read_csv("aisles.csv")

order_product.head()

products.head()

orders.head()

aisles.head()

数据基本处理

# 合并表格
table1 = pd.merge(order_product,products,on=["product_id","product_id"])
table2 = pd.merge(table1,orders,on=["order_id","order_id"])
table = pd.merge(table2,aisles,on=["aisle_id","aisle_id"])

tabel.shape

table.head()

# 交叉表合并
data = pd.crosstab(table["user_id"],table["aisle"])
data.shape

data.head()

# 数据截取
new_data = data[:1000]

特征工程-pca

transfer = PCA(n_components=0.9)
trans_data = transfer.fit_transform(new_data)
trans_data.shape

机器学习（k-means）

estimator = KMeans(n_clusters=5)
pre_data = estimator.fit_predict(trans_data)

模型评估

silhouette_score(trans_data,pre_data)