1.演算法簡介

kNN的思想很簡單：計算待分類的資料點與訓練集所有樣本點，取距離最近的k個樣本；統計這k個樣本的類別數量；根據多數表決方案，取數量最多的那一類作為待測樣本的類別。距離度量可採用Euclidean distance，Manhattan distance和cosine。

用Iris資料集作為測試，程式碼參考[1]

import numpy as np
import scipy.spatial.distance as ssd

def read_data(fn):
    """ read dataset and separate into characteristics data
        and label data
    """
 
    # read dataset file
    with open(fn) as f:
        raw_data = np.loadtxt(f, delimiter= ',', dtype="float", 
            skiprows=1, usecols=None)

    #initialize
    charac=[]; label=[]
    #obtain input characrisitics and label
    for row in raw_data:
        charac.append(row[:-1])
        label.append(int (row[-1]))
    return np.array(charac),np.array(label)

def knn(k,dtrain,dtest,dtr_label):
    """k-nearest neighbors algorithm"""

    pred_label=[]
    #for each instance in test dataset, calculate
    #distance in respect to train dataset
    for di in dtest:
        distances=[]
        for ij,dj in enumerate(dtrain):
            distances.append((ssd.euclidean(di,dj),ij))

        #sort the distances to get k-neighbors
        k_nn=sorted(distances)[:k]

        #classify accroding to the maxmium label
        dlabel=[]
        for dis,idtr in k_nn:
            dlabel.append(dtr_label[idtr])
        pred_label.append(np.argmax(np.bincount(dlabel)))

    return pred_label

def evaluate(result):
    """evaluate the predicited label"""

    eval_result=np.zeros(2,int)
    for x in result:
        #pred_label==dte_label
        if x==0:
            eval_result[0]+=1
        #pred_label!=dte_label
        else:
            eval_result[1]+=1

    return eval_result


dtrain,dtr_label=read_data('iris-train.csv')
dtest,dte_label=read_data('iris-test.csv')

K=[1,3,7,11]

print "knn classification result for iris data set:\n"
print "k    | number of correct/wrong classified test records"

for k in K:
    pred_label=knn(k,dtrain,dtest,dtr_label)
    eval_result=evaluate(pred_label-dte_label)

    #print the evaluted result into screen
    print k,"   | ", eval_result[0], "/", eval_result[1]

print
 

2. Referrence