AdaBoost.py

#encoding: utf-8
'''
    AdaBoost算法，配合决策树桩
'''
import sys
from os import path
import numpy as np
import matplotlib.pyplot as plt
import matplotlib.patches as patches
from functools import reduce
sys.path.append(path.dirname(path.dirname(path.abspath(__file__))))
from dataSet.watermelon_3alpha import wm_dataSet, watermelon_counterexample_x, watermelon_posiexam_x
from DecisionTreeStumps_Model import generate_stumps

TOTAL_TURNS = 8
POSI_COLOR = "#DAF9CA"
COUNTER_COLOR = "#C7B3E5"
def plot_init():
    '''
    初始化matplot，并将样本点绘制到ax中
    '''
    ax1 = plt.subplots()[1]
    ax1.set_title(u'AdaBoost')
    ax1.set_xlabel(u"Density")
    ax1.set_ylabel(u"Sugar Rate")
    ax1.set_xlim(0, 1)
    ax1.set_ylim(0, 0.7)
    return ax1

def main():
    '''
        主函数
    '''
    #1 初始化权值和轮次
    weights = [1/len(wm_dataSet)] * len(wm_dataSet)
    #2 AdaBoost
    def boosting(turns, packages, weights):
        '''
            尾递归实现boosting
            输入:turns: 剩余轮次，
            packages(权值alpha，分类器，[错误率，分界点，分界策略,属性下标])，weights：当前权值
            输出 packages
        '''
        if not turns:
            def final_classifier(feature):
                '''
                    最终的组合分类器
                '''
                result = sum(map(lambda item: item[0]*item[1](feature), packages))
                return 1 if result > 0 else -1
            return packages, final_classifier

        classifier, pack = generate_stumps(wm_dataSet, weights)
        error = pack[0]
        if error == 0:
            return
        alpha = 0.5 * np.log((1-error)/error) #分类器权重
        weights = list(map(lambda sample, weigh: weigh*np.exp(-alpha*sample[1]*classifier(sample[0])), wm_dataSet, weights))
        factor = sum(weights)#归一化因子
        weights = list(map(lambda weigh: weigh/factor, weights))
        packages.append((alpha, classifier, pack))
        return  boosting(turns-1, packages, weights)

    detail_info_package, adaBoost_classifier = boosting(TOTAL_TURNS, [], weights)
    print(list(map(adaBoost_classifier, watermelon_counterexample_x+watermelon_posiexam_x)))

    #3 求各个边界
    ax = plot_init()
    x_edges = [0, 1]
    y_edges = [0, 1]
    for var in detail_info_package:
        if var[2][-1] == 0:
            x_edges.append(var[2][1])
            ax.plot([x_edges, x_edges], [0, 1])
        else:
            y_edges.append(var[2][1]) 
            ax.plot([0, 1], [y_edges, y_edges])
    x_edges = sorted(list(set(x_edges)))
    y_edges = sorted(list(set(y_edges)))
    '''
        去除重复边界并排序
    '''


    #4 求边界构成的最小矩形并且填色
    def y_func(y_pre_edge, y_post_edge): 
         return list(map(lambda x_pre_edge, x_post_edge: ([x_pre_edge, y_pre_edge], [x_post_edge-x_pre_edge, y_post_edge-y_pre_edge]), x_edges, x_edges[1: ]))
    rects = list(reduce(lambda l1, l2: l1+l2, map(y_func, y_edges, y_edges[1: ])))
    for rect in rects:
        color = POSI_COLOR if adaBoost_classifier([rect[0][0] + rect[1][0]/2, rect[0][1] + rect[1][1]/2]) == 1 else COUNTER_COLOR
        ax.add_patch(
            patches.Rectangle(tuple(rect[0]), rect[1][0], rect[1][1], color=color),
        )

    #5 绘制样本点并且绘制边界线（edges）
    pose_points, counter_points = zip(*watermelon_posiexam_x), zip(*watermelon_counterexample_x)
    ax.plot(next(pose_points), next(pose_points), 'go')
    ax.plot(next(counter_points), next(counter_points), 'bo')#样本点
    plt.show()
if __name__ == '__main__':
    main()