KNN算法實現

• 前提基礎：

  KNN基本原理，本文參考 李航博士 著《統計學習方法》

• 距離度量方式

1. 歐式距離
2. 曼哈頓距離
3. 夾角餘弦
4. 切比雪夫距離
5. 馬氏距離

• k值的選擇

 

• 算法優缺點

1. 優點：精度高、對異常值不敏感、無數據輸入假定；
2. 缺點：計算複雜度高、空間複雜度高；

    適用數據範圍：數值型和標稱型 

• 代碼如下：

#encoding=utf-8
import pandas as pd
import numpy as np
import time

from sklearn.cross_validation import train_test_split
from sklearn.metrics import accuracy_score

class KNN(object):

    def __init__(self):
        self.k = 10

    def predict(self, testset, trainset, train_labels):
        predict_ = []
        count = 0
        for test_vec in testset:
            print count
            count += 1
            knn_list = []    #當前k個最近鄰居
            max_index = -1   #當前k個最近鄰居中距離最遠點的座標
            max_dist = 0     #當前k個最近鄰居中距離最遠點的距離

            # 先將前k個點放入k個最近鄰居中，充滿knn_list
            for i in xrange(self.k):
                label = train_labels[i]
                train_vec = trainset[i]

                dist = np.linalg.norm(train_vec - test_vec)

                knn_list.append((dist, label))

            # 剩下的點
            for i in xrange(self.k, len(train_labels)):
                label = train_labels[i]
                train_vec = trainset[i]

                dist = np.linalg.norm(train_vec - test_vec)

                # 尋找k個鄰近點距離最遠的點
                if max_index < 0:
                    for j in xrange(10):
                        if max_dist < knn_list[j][0]:
                            max_index = j
                            max_dist = knn_list[max_index][0]

                if dist < max_dist:
                    knn_list[max_index] = (dist, label)
                    max_index = -1
                    max_dist = 0

            class_total = 10
            class_count = [0 for i in xrange(class_total)]
            for dist, label in knn_list:
                class_count[label] += 1

            mmax = max(class_count)

            for i in xrange(class_total):
                if mmax == class_count[i]:
                    predict_.append(i)
                    break
        return np.array(predict_)




if __name__ == '__main__':

    print 'Start read data'

    time1 = time.time()

    raw_data = pd.read_csv("../data/train.csv", header=0)
    time2 = time.time()
    print 'Read data cost ', time2 - time1, ' second', '\n'

    print raw_data.info()
    print raw_data.head()

    data = raw_data.values

    imgs = data[0::, 1::]
    labels = data[::, 0]

    train_features, test_features, train_labels, test_labels = train_test_split(
        imgs, labels, test_size=0.25, random_state=33)

    print train_features.shape
    print test_features.shape

    print 'Start predicting'
    k = KNN()
    test_predict = k.predict(test_features, train_features, train_labels)
    time3 = time.time()
    print 'Predicting cost ', time3 - time2, ' seocnd', '\n'

    score = accuracy_score(test_labels, test_predict)
    print 'The accuracy score is ', score