DBSCAN 将远处的点聚集在一起

Question

想要计算每个集群内的 Instagram/酒店比率。但结果表明我实际上将非常遥远的点聚集在一起。 DBSCAN 不太可能出现这种情况。怎么了？

程序：使用 DBSCAN 对 instagram 帖子进行聚类，然后使用 1NN 对酒店进行分类。

import pandas as pd
import matplotlib.pyplot as plt
import numpy as np
from sklearn.cluster import KMeans
from sklearn.cluster import DBSCAN
from sklearn.neighbors import KNeighborsClassifier


if __name__ == '__main__':
    inst = pd.read_csv('inst.csv', encoding='utf-8')
    ht = pd.read_csv('ht.csv', encoding='utf-8')

    inst = inst[(inst.lat >= 48.30) & (inst.lng >= -139.06) & (inst.lat <= 60.00) & (inst.lng <= -114.03)]
    ht = ht[(ht.lat >= 48.30) & (ht.lng >= -139.06) & (ht.lat <= 60.00) & (ht.lng <= -114.03)]

    # kmean = KMeans(n_clusters=50,n_jobs=-1)
    # kmean.fit(inst[['lat', 'lng']])
    #
    # ht_labels = kmean.predict(ht[['lat', 'lng']])
    # inst_labels = kmean.predict(inst[['lat', 'lng']])
    #
    # plt.scatter(ht.lng, ht.lat, c=ht_labels, alpha=0.5)
    # plt.savefig('./fig/hotel_clusters.png')
    # plt.clf()
    #
    # plt.scatter(inst.lng, inst.lat, c=inst_labels, alpha=0.5)
    # plt.savefig('./fig/instagram_posts_clusters.png')
    # plt.clf()

    dbs = DBSCAN(eps=0.05,min_samples=10,metric='haversine', n_jobs=-1)

    ht_labels = dbs.fit_predict(ht[['lat', 'lng']])
    inst_labels = dbs.fit_predict(inst[['lat', 'lng']])

    plt.scatter(ht.lng, ht.lat, c=ht_labels, alpha=0.5)
    plt.savefig('./fig/hotel_clusters1.png')
    plt.clf()

    plt.scatter(inst.lng, inst.lat, c=inst_labels, alpha=0.5)
    plt.savefig('./fig/instagram_posts_clusters1.png')
    plt.clf()

    knn = KNeighborsClassifier(n_neighbors=1, n_jobs=-1)

    knn.fit(inst[['lat', 'lng']], inst_labels)
    ht_labels = knn.predict(ht[['lat', 'lng']])

    plt.scatter(ht.lng, ht.lat, c=ht_labels, alpha=0.5)
    plt.savefig('./fig/hotel_clusters3.png')
    plt.clf()

    plt.scatter(inst.lng, inst.lat, c=inst_labels, alpha=0.5)
    plt.savefig('./fig/instagram_posts_clusters3.png')
    plt.clf()

    ht = ht[['lat', 'lng']]
    ht['lb'] = ht_labels
    inst = inst[['lat', 'lng']]
    inst['lb'] = inst_labels

    ht1 = ht.groupby(['lb']).count().reset_index().set_index('lb')
    inst1 = inst.groupby(['lb']).count().reset_index().set_index('lb')

    print(ht1)
    ratio = inst1/ht1
    print(ratio)

    clu = 2
    plt.scatter(ht[ht.lb == clu].lng, ht[ht.lb == clu].lat, c='black')
    print(len(ht[ht.lb == clu]))
    plt.scatter(ht.lng, ht.lat, c=ht_labels, alpha=0.05)
    plt.savefig('./fig/hotel_clusters4.png')
    plt.clf()

我找不到错误，你能帮忙吗？

Answer 1

Sklearn havrsine 需要以弧度为单位的坐标，而不是度数。

所以我认为 mod 2pi 头点很接近。