knn算法实现

knn.py

本文为 落魄陶陶 原创,转载请注明出处
数据来源及源码参见github

• 学习并参考《机器学习实战》第二章
• 主要使用Pandas库
• knn.py为基本算法实现,基于数据knn.xlsx
• knn_dating.py为约会预测,基于数据datingTestSet.txt

---

knn.py

    # 1.读取数据 得到 每行都是 x1,x2,x3,...,xn,y的dataFrame df
    # 2.计算给定点target = (x1,x2,...,xn)与步骤1中df的所有点的距离
    # 3.对步骤2所得的所有距离排序 asc,取前n个对应的y
    # 4.步骤3中y出现频率最高的为target的y

from collections import namedtuple

import numpy as np
import pandas as pd

MinMaxDict = namedtuple('MinMax', 'min max range')


def load_data()-> pd.DataFrame:
    """
    加载数据
    :return: pd.DataFrame 最后一列为label,其余为输入
    """
    return pd.read_excel('knn.xlsx')


def norm(df):
    min_max_list = []
    for column in df.columns[:-1]:
        col = df[column]
        min_val, max_val = col.min(), col.max()
        length = max_val - min_val
        df[column] = (col - min_val) / length
        min_max_list.append(MinMaxDict(min_val, max_val, length))
    return df, min_max_list


def norm_target(target, min_max_list):
    ret = []
    for i, item in enumerate(min_max_list):
        ret.append((target[i] - item.min) / item.range)
    return tuple(ret)


def classify(n, df, target):
    """
    构建一个DataFrame,包含[labels,distance]两列,
    label就是训练集的label,distance为目标数据targe对训练集df每条数据的欧氏距离
    然后再根据distance排序,取前n个结果
    再对前n个结果统计,个数最多的label就是分类结果
    :param n:
    :param df:
    :param target:
    :return:
    """
    label_column = df.columns[-1]
    result = pd.DataFrame()
    result['labels'] = df[label_column].copy()
    result['distance'] = np.sqrt(np.sum(df.loc[:, df.columns[:-1]].sub(target, axis='columns') ** 2, axis=1))
    result = result.sort_values(by='distance', ascending=True)
    return result[:n]['labels'].value_counts(ascending=False).index[0]


if __name__ == '__main__':
    n = 5
    # 1.读取数据得到DataFrame
    df = load_data()
    # 2.将数据正规化,同时得到每列最大最小及跨度信息,用于目标数据正规化
    x_columns = df.columns[:-1]
    df, min_max_list = norm(df)
    # 3.将目标正规化
    item = (0, 10)
    item = norm_target(item, min_max_list)
    # 4.对目标值和已有数据计算欧氏距离,取前n个最小值得到分类结果
    result = classify(n, df, item)
    print(result)

---

约会预测
knn_dating.py

import pandas as pd

import knn


def load_dating_data(file, ratio):
    df = pd.read_csv(file, sep='\t')
    df.columns = ['fly_miles', 'game_time', 'ice_cream', 'labels']
    df.labels = df.labels.map(lambda x: ['smallDoses', 'didntLike', 'largeDoses'].index(x) + 1)
    size = df.shape[0]
    num = int(size * ratio)
    return df[:num], df[num:]


# class DatingKnn(KNN):
#
#     def __init__(self, n, ratio):
#         super().__init__(n)
#         num = int(self.df.shape[0] * ratio)
#         self.testing_set = self.df[num:]
#         self.df = self.df[:num]
#
#     def load_data(self):
#         return load_dating_data('datingTestSet.txt')
#
#     def test(self):
#         results = pd.Series([self.classify(item[1:-1], False) for item in self.testing_set.itertuples()])
#         labels = self.testing_set.labels
#         labels.index = results.index
#         bingo = (labels - results).value_counts()[0]
#         print(bingo / results.shape[0])
#         # print(labels-results)
#         # print(self.testing_set.labels)


if __name__ == '__main__':
    # df = load_dating_data('datingTestSet.txt')
    # print(df.labels.unique())
    # fig = plt.figure()
    # ax = fig.add_subplot(111)
    # ax.scatter(df.iloc[:, 0], df.iloc[:, 1], s=15 * df.labels, c=15 * df.labels)
    # plt.show()

    n = 5
    ratio = 0.9  # 数据的百分之90为训练数据,10%为测试数据
    # 1.读取数据得到DataFrame
    training_set, testing_set = load_dating_data('datingTestSet.txt', 0.9)
    # 2.将数据正规化,同时得到每列最大最小及跨度信息,用于目标数据正规化
    df, min_max_list = knn.norm(training_set)

    results_list = []
    for item in testing_set.itertuples(index=False):
        # 3.将目标正规化
        item = item[:-1]
        item = knn.norm_target(item, min_max_list)
        # 4.对目标值和已有数据计算欧氏距离,取前n个最小值得到分类结果
        result = knn.classify(n, df, item)
        results_list.append(result)

    # 5.使用测试集计算正确率
    total = len(results_list)
    testing_set.reset_index(inplace=True)
    labels = testing_set[testing_set.columns[-1]] # 两个Series相减时,按照index相同的相减,所以重置测试集的index
    results = pd.Series(results_list)
    bingo = (labels - results_list).value_counts()[0] # 对测试集label和预测结果做差,统计结果为0的个数即为正确的个数
    percent = bingo / total
    print(percent)