7-8. DBSCAN

DBSCAN 적용하기 – 붓꽃 데이터 셋

from sklearn.datasets import load_iris

import matplotlib.pyplot as plt
import numpy as np
import pandas as pd
%matplotlib inline

iris = load_iris()
feature_names = ['sepal_length','sepal_width','petal_length','petal_width']

# 보다 편리한 데이타 Handling을 위해 DataFrame으로 변환
irisDF = pd.DataFrame(data=iris.data, columns=feature_names)
irisDF['target'] = iris.target
irisDF.head()



	sepal_length	sepal_width	petal_length	petal_width	target
0	5.1	3.5	1.4	0.2	0
1	4.9	3.0	1.4	0.2	0
2	4.7	3.2	1.3	0.2	0
3	4.6	3.1	1.5	0.2	0
4	5.0	3.6	1.4	0.2	0

 

eps 0.6 min_samples=8 로 DBSCAN 군집화 적용

from sklearn.cluster import DBSCAN

dbscan = DBSCAN(eps=0.6, min_samples=8, metric='euclidean')
dbscan_labels = dbscan.fit_predict(iris.data)

irisDF['dbscan_cluster'] = dbscan_labels

iris_result = irisDF.groupby(['target'])['dbscan_cluster'].value_counts()
print(iris_result)



target  dbscan_cluster
0        0                49
        -1                 1
1        1                46
        -1                 4
2        1                42
        -1                 8
Name: dbscan_cluster, dtype: int64

 

### 클러스터 결과를 담은 DataFrame과 사이킷런의 Cluster 객체등을 인자로 받아 클러스터링 결과를 시각화하는 함수  
def visualize_cluster_plot(clusterobj, dataframe, label_name, iscenter=True):
    if iscenter :
        centers = clusterobj.cluster_centers_
        
    unique_labels = np.unique(dataframe[label_name].values)
    markers=['o', 's', '^', 'x', '*']
    isNoise=False

    for label in unique_labels:
        label_cluster = dataframe[dataframe[label_name]==label]
        if label == -1:
            cluster_legend = 'Noise'
            isNoise=True
        else :
            cluster_legend = 'Cluster '+str(label)
        
        plt.scatter(x=label_cluster['ftr1'], y=label_cluster['ftr2'], s=70,\
                    edgecolor='k', marker=markers[label], label=cluster_legend)
        
        if iscenter:
            center_x_y = centers[label]
            plt.scatter(x=center_x_y[0], y=center_x_y[1], s=250, color='white',
                        alpha=0.9, edgecolor='k', marker=markers[label])
            plt.scatter(x=center_x_y[0], y=center_x_y[1], s=70, color='k',\
                        edgecolor='k', marker='$%d$' % label)
    if isNoise:
        legend_loc='upper center'
    else: legend_loc='upper right'
    
    plt.legend(loc=legend_loc)
    plt.show()

 

PCA 2개 컴포넌트로 기존 feature들을 차원 축소 후 시각화

from sklearn.decomposition import PCA
# 2차원으로 시각화하기 위해 PCA n_componets=2로 피처 데이터 세트 변환
pca = PCA(n_components=2, random_state=0)
pca_transformed = pca.fit_transform(iris.data)
# visualize_cluster_2d( ) 함수는 ftr1, ftr2 컬럼을 좌표에 표현하므로 PCA 변환값을 해당 컬럼으로 생성
irisDF['ftr1'] = pca_transformed[:,0]
irisDF['ftr2'] = pca_transformed[:,1]

visualize_cluster_plot(dbscan, irisDF, 'dbscan_cluster', iscenter=False)

 

eps의 크기를 증가 한 후 노이즈 확인

from sklearn.cluster import DBSCAN

dbscan = DBSCAN(eps=0.8, min_samples=8, metric='euclidean')
dbscan_labels = dbscan.fit_predict(iris.data)

irisDF['dbscan_cluster'] = dbscan_labels
irisDF['target'] = iris.target

iris_result = irisDF.groupby(['target'])['dbscan_cluster'].value_counts()
print(iris_result)

visualize_cluster_plot(dbscan, irisDF, 'dbscan_cluster', iscenter=False)


target  dbscan_cluster
0        0                50
1        1                50
2        1                47
        -1                 3
Name: dbscan_cluster, dtype: int64

 

min_samples의 크기를 증가 후 노이즈 확인

dbscan = DBSCAN(eps=0.6, min_samples=16, metric='euclidean')
dbscan_labels = dbscan.fit_predict(iris.data)

irisDF['dbscan_cluster'] = dbscan_labels
irisDF['target'] = iris.target

iris_result = irisDF.groupby(['target'])['dbscan_cluster'].value_counts()
print(iris_result)
visualize_cluster_plot(dbscan, irisDF, 'dbscan_cluster', iscenter=False)




target  dbscan_cluster
0        0                48
        -1                 2
1        1                44
        -1                 6
2        1                36
        -1                14
Name: dbscan_cluster, dtype: int64