【代码复现】MIDIE

1、算法

本文主要是论文《Two-stage instance selection and adaptive bag mapping algorithm for multi-instance learning》中算法代码的复现。具体算法原理见此文。本文与SMDP有类似之处，都使用了DP聚类。

2、代码

2.1 DIP.py

这一部分代码是创建实例原型池，即：找出所有包的代表实例聚集在一起。此阶段为在包内选取实例原型。

import math
import numpy as np
from scipy.spatial.distance import cdist
from MIL_MIDIE.MIL_Frame.MILTool import get_cosineclass DIP:#基于DIP(discriminative instance pool)映射def __init__(self,bags,scale_num):self.bags=bagsself.scale_num=scale_numself.inner_bag_distance=[]self.discriminative_instance = self.__get_discriminative_ins()def __get_discriminative_ins(self):#获得discriminative实例"""Get the discriminative instance of the each bags to build the discriminative instance pool.:return:"""for bag_i in range(self.bags.shape[0]):self.__get_instance_in_bag(self.bags[bag_i])return np.array(self.inner_bag_distance)def __get_instance_in_bag(self, bag):#从包中找到实例作为discriminative instance# Step 1. Calculate the distances of each instance.计算实例间的距离# print(bag[0].shape[0])ins_space_to_bag = []for ins in bag[0][:, :-1]:ins_space_to_bag.append(ins)ins_space_to_bag = np.array(ins_space_to_bag)distance_ins_to_ins = cdist(ins_space_to_bag, ins_space_to_bag)#计算两个实例间的马氏距离# Step 2. Calculate the affinity of each instance.计算实例间的相似度，通过计算余弦值来进行度量affinity_ins_to_ins = get_cosine(ins_space_to_bag)affinity_ins = np.zeros((len(distance_ins_to_ins), len(distance_ins_to_ins))).astype("int32")affinity_ins_score = []# density_ins_score = []ave_dis_ins = affinity_ins_to_ins.mean()for i in range(len(affinity_ins_to_ins)):for j in range(len(affinity_ins_to_ins)):if affinity_ins_to_ins[i, j] <= ave_dis_ins:affinity_ins[i, j] = 1affinity_ins_score.append(sum(affinity_ins[i]))# Step 3. Calculate the density of each instance.dis_cut = 0.4 * distance_ins_to_ins.max()density_ins_score = np.zeros(len(distance_ins_to_ins)).astype("float64")for i in range(len(distance_ins_to_ins)):if dis_cut == 0:density_ins_score[i] = 1else:density_ins_score[i] = sum(np.exp(-(distance_ins_to_ins[i] / dis_cut) ** 2))# Step 4. Calculate the lambda of each instance.lambda_ins_score = np.multiply(affinity_ins_score, density_ins_score).tolist()# Step 5. Get the discriminative instance.for i in range(math.ceil(self.scale_num * bag[0].shape[0])):self.inner_bag_distance.append(ins_space_to_bag[lambda_ins_score.index(min(lambda_ins_score))])lambda_ins_score[lambda_ins_score.index(min((lambda_ins_score)))] = -1if __name__=='__main__':a = np.array([0, 1, 3])b = np.array([[12, 3, 4], [0, 1, 2], [2, 5, 6], [1, 2, 3]])c=cdist([a], b)print(min(max(c)))

2.2 SDI.py

这段代码主要是进行第二阶段：在实例原型池中选出代表实例。先要通过DIP得到实例原型池，再通过DP来选出代表实例。在这个过程中，需要计算实例原型的密度 $\rho$ 以及实例到比他实例密度更大的实例的距离 $δ$ 。最后计算 $\lambda$ 来找出代表实例。

import numpy as np
from scipy.spatial.distance import cdist
import MIL_MIDIE.MIDIE.DIP as DIP
from MIL_MIDIE.MIL_Frame.MILTool import MILToolclass SDI:def __init__(self, bags, ratio_instance_to_bag, num_SDI):self.bags = bagsself.ratio_ins = ratio_instance_to_bagself.num_SDI = num_SDIself.discriminative_instance = self.__print_SDI()self.final_discriminative_instance = self.__select_discriminative_instance()def __print_SDI(self):DIP_demo = DIP.DIP(self.bags,self.ratio_ins)#输入包以及搜索范围return DIP_demo.discriminative_instance#返回找到的discriminative instancedef __select_discriminative_instance(self):#通过DP来选出代表实例# Step 1. Calculate the distance of discriminative instances.discriminative_instance_distance = cdist(self.discriminative_instance, self.discriminative_instance)# Step 2. The cutoff distances.dis_cut = 0.4 * discriminative_instance_distance.max()# The density of  discriminative instances.density_discriminative_ins = np.zeros(len(discriminative_instance_distance)).astype("float64")for i in range(len(discriminative_instance_distance)):if dis_cut == 0:density_discriminative_ins[i] = 1else:density_discriminative_ins[i] = sum(np.exp(-(discriminative_instance_distance[i] / dis_cut) ** 2))# Step 3. The distance of closest instance that is denser than itself.distance_closest = []for i in range(len(density_discriminative_ins)):more_density_instance_index = []temp_density_instance = density_discriminative_ins[i]for j in range(len(density_discriminative_ins)):if density_discriminative_ins[j] > temp_density_instance:more_density_instance_index.append(j)temp_distance_more_instance = []for index in range(0, len(more_density_instance_index)):index_k = more_density_instance_index[index]temp_distance_more_instance.append(discriminative_instance_distance[i][index_k])if temp_distance_more_instance:temp_distance_more_instance.sort()distance_closest.append(temp_distance_more_instance[0])else:distance_closest.append(float('inf'))# Step 4. The lambda of discriminative instance.lambda_discriminative_instance = np.multiply(distance_closest, density_discriminative_ins).tolist()final_discriminative_instance = []for i in range(self.num_SDI):index_most = lambda_discriminative_instance.index(max(lambda_discriminative_instance))final_discriminative_instance.append(self.discriminative_instance[index_most])lambda_discriminative_instance[index_most] = -1return np.array(final_discriminative_instance)if __name__ == '__main__':file_path = "MUSK1_1.arff"mil = MILTool(file_path)bags = mil.bagsSDI_demo = SDI(bags, 0.01, 2)ins = SDI_demo.final_discriminative_instanceprint(type(ins))

2.3 DIE.py

这一部分的代码主要做的就是包映射相关工作。可以对不同状态的包进行包获取，包状态主要有三种：global、positive、negative。也可以通过相加（add）或拼接（con）两种方式来对包进行映射。最后对向量进行归一化。

import warnings
import numpy as np
from MIL_MIDIE.MIDIE.SDI import SDI
from MIL_MIDIE.MIL_Frame.MILTool import MILTool, dis_euclidean, get_ten_fold_indexwarnings.filterwarnings('ignore')class DIE:def __init__(self, all_bag, tr_index, bags_status, embed_status, ratio_instance_to_bag,num_discriminative_instance):self.bags = all_bagself.bags_status = bags_statusself.embed_status = embed_statusself.tr_index = tr_indexself.ra_ins = ratio_instance_to_bagself.num_dis_ins = num_discriminative_instanceself.train_final_bag = self.__get_bags()self.embedding_vector = self.__embedding()def __get_bags(self):"""Get train bags through three types:g: The source of bags is the global bags.p: The source of bags is the positive bags.n：The source of bags is the negative bags:return: The bags."""if self.bags_status == 'g':  # global bagsreturn self.bags[self.tr_index]  # 训练集索引的所有包elif self.bags_status == 'p':  # positive bagspositive_bags_index = []  # 正包for i in range(len(self.tr_index)):if self.bags[self.tr_index[i], -1] == 1:  # 标签为1positive_bags_index.append(self.tr_index[i])  # 加入矩阵中return self.bags[positive_bags_index]elif self.bags_status == 'n':  # negative bagsnegative_bags_index = []  # 负包for i in range(len(self.tr_index)):if not self.bags[self.tr_index[i], -1] == 1:  # 标签不为1negative_bags_index.append(self.tr_index[i])return self.bags[negative_bags_index]def __embedding(self):"""Let a bag embedding into a single vector.:return:"""# Step 1. Get the discriminative instance through the select discriminative instance method./通过SDI方法找出代表实例discriminative_instance = SDI(self.train_final_bag, self.ra_ins, self.num_dis_ins).final_discriminative_instance# Step 2. Embedding the bags as the single vector./将包映射为向量bag_to_vector = []for bag_i in range(self.bags.shape[0]):# Step 2.1 Find the embedding model./选择嵌入模型temp_single_vector = []if self.embed_status == 'add':#映射方法为相加temp_single_vector = np.zeros(self.bags[bag_i][0].shape[1] - 1).astype("float64")elif self.embed_status == 'con':#映射方法为拼接temp_single_vector = np.zeros(self.num_dis_ins * (self.bags[bag_i][0].shape[1] - 1)).astype("float64")else:print('Your input model is not exist!\n')break# Step 2.2 Specific steps of embedding.for ins_i in self.bags[bag_i][0][:, :-1]:#所有实例进行寻找temp_distance_dis_to_ins = []# Step 2.2.1 Find  the nearest discriminative instance.#找到最近的代表实例for dis_ins_i in range(self.num_dis_ins):#与所有的代表进行距离计算temp_distance_dis_to_ins.append(dis_euclidean(ins_i, discriminative_instance[dis_ins_i]))#计算欧式距离temp_index = temp_distance_dis_to_ins.index(min(temp_distance_dis_to_ins))#将距离最小的那个索引纳入# Subtract the i-th instance and the nearest discriminative instance.temp_dis_to_ins_vector = ins_i - discriminative_instance[temp_index]if self.embed_status == 'add':temp_single_vector += temp_dis_to_ins_vectorelif self.embed_status == 'con':start_index = 0end_index = self.bags[bag_i][0].shape[1] - 1temp_single_vector[start_index + temp_index * end_index:(temp_index + 1) * end_index] += (temp_dis_to_ins_vector)# Step 3. Normalize the vectortemp_single_vector = np.sign(temp_single_vector) * np.sqrt(np.abs(temp_single_vector))temp_norm = np.linalg.norm(temp_single_vector)temp_single_vector = temp_single_vector / temp_normbag_to_vector.append(temp_single_vector)return np.array(bag_to_vector)if __name__ == '__main__':file_path = "rec_sport_hockey.mat"mil = MILTool(file_path)bags = mil.bagstrain_index, te_index = get_ten_fold_index(bags)miDie_demo = DIE(bags, train_index[1], 'g', 'con', 0.01, 2).embedding_vectorprint(miDie_demo)miDie_demo = np.array(miDie_demo)

2.4 DIEPredict.py

这一部分主要是对训练集进行预测，并且分别使用KNN、决策树、线性支持向量机、高斯径向基函数支持向量机对处理后的数据集进行分类预测，最后输出预测精度。

import warnings
import numpy as np
from sklearn.metrics import f1_score,accuracy_score,roc_auc_score
from MIL_MIDIE.MIDIE.DIE import DIE
from MIL_MIDIE.MIL_Frame.MILTool import MILTool
from MIL_MIDIE.MIL_Frame.MILTool import get_ten_fold_indexwarnings.filterwarnings('ignore')class DIEPredict:def __init__(self, file_path, bags_status, embed_status, ratio_instance_to_bag,num_discriminative_instance, bags=None):''':param file_path: 数据文件路径:param bags_status: 包状态，即正包、负包、所有包:param embed_status: 映射方法，有相加（add）或拼接（con）:param ratio_instance_to_bag: 包中实例原型选择比例:param num_discriminative_instance:代表实例数量:param bags:包'''self.file_path = file_pathself.bags_status = bags_statusself.embed_status = embed_statusself.ra_ins = ratio_instance_to_bagself.num_dis_ins = num_discriminative_instanceself.bags = bagsself.__DIEPredict()def __DIEPredict(self):MIL = MILTool(para_file_name=self.file_path, bags=self.bags)#加载数据集bags = MIL.bags#加载包bags_label = MIL.bags_label#加载标签num_fold = 10accuracy_res_score = np.zeros(4).astype('float64')#预测精确度分数f1_res_score = np.zeros(4).astype('float64')#f1分数来评价预测精度roc_aoc_res_score = np.zeros(4).astype('float64')accuracy_std = [[], [], [], []]f1_res_std = [[], [], [], []]roc_aoc_res_std = [[], [], [], []]for i in range(num_fold):#四种分类器knn_estimator = MIL.get_classifier('knn')#k近邻DTree_estimator = MIL.get_classifier('DTree')#决策树LSVM_estimator = MIL.get_classifier('linear_svm')#线性支持向量机RSVM_estimator = MIL.get_classifier('rbf_svm')#高斯径向基函数核accuracy_measure_score = np.zeros(4).astype('float64')fl_measure_score = np.zeros(4).astype('float64')roc_measure_score = np.zeros(4).astype('float64')tr_index, te_index = get_ten_fold_index(bags)temp_f1_score = 10temp_acc_score = 10temp_roc_score = 10for j in range(10):tr_bags = bags[tr_index[j]]#本次训练包tr_label_true = bags_label[[tr_index[j]]]#训练包标签te_bags = bags[te_index[j]]#本次测试包te_label_true = bags_label[[te_index[j]]]#测试包标签#第1步 使用算法将包映射test_Demo = DIE(bags, tr_index[j], self.bags_status, self.embed_status, self.ra_ins,self.num_dis_ins).embedding_vectortr_embedding_vector = test_Demo[tr_index[j]]te_embedding_vector = test_Demo[te_index[j]]#第2.1.1步 使用knn进行分类knn_model = knn_estimator.fit(tr_embedding_vector, tr_label_true)#输入映射后的向量以及标签进行处理te_label_predict_knn = knn_model.predict(te_embedding_vector)# 2.1.2 计算三种预测分数fl_measure_score[0] += f1_score(te_label_true, te_label_predict_knn)#将测试集标签与测试集预测结果进行评分accuracy_measure_score[0] += accuracy_score(te_label_true, te_label_predict_knn)try:temp_aoc = roc_auc_score(te_label_true, te_label_predict_knn)roc_measure_score[0] += temp_aocexcept ValueError:pass#第2.2.1步 使用决策树进行分类DTree_model = DTree_estimator.fit(tr_embedding_vector, tr_label_true)te_label_predict_tree = DTree_model.predict(te_embedding_vector)#2.2.2 计算三种预测分数fl_measure_score[1] += f1_score(te_label_true, te_label_predict_tree)accuracy_measure_score[1] += accuracy_score(te_label_true, te_label_predict_tree)try:temp_aoc = roc_auc_score(te_label_true, te_label_predict_knn)roc_measure_score[1] += temp_aocexcept ValueError:pass#第2.3.1步 使用线性支持向量机LSVM_model = LSVM_estimator.fit(tr_embedding_vector, tr_label_true)te_label_predict_LSVM = LSVM_model.predict(te_embedding_vector)#2.3.2 计算三种预测分数fl_measure_score[2] += f1_score(te_label_true, te_label_predict_LSVM)accuracy_measure_score[2] += accuracy_score(te_label_true, te_label_predict_LSVM)try:temp_aoc = roc_auc_score(te_label_true, te_label_predict_knn)roc_measure_score[2] += temp_aocexcept ValueError:pass#第2.4.1 使用高斯径向基函数核支持向量机RSVM_model = RSVM_estimator.fit(tr_embedding_vector, tr_label_true)te_label_predict_RSVM = RSVM_model.predict(te_embedding_vector)#2.4.2 计算三种预测分数fl_measure_score[3] += f1_score(te_label_true, te_label_predict_RSVM)accuracy_measure_score[3] += accuracy_score(te_label_true, te_label_predict_RSVM)try:temp_aoc = roc_auc_score(te_label_true, te_label_predict_knn)roc_measure_score[3] += temp_aocexcept ValueError:pass#计算本次训练的准确度accuracy_res_score[0] += accuracy_measure_score[0] / temp_acc_score  # Knn   accaccuracy_res_score[1] += accuracy_measure_score[1] / temp_acc_score  # DTree accaccuracy_res_score[2] += accuracy_measure_score[2] / temp_acc_score  # LSVM  accaccuracy_res_score[3] += accuracy_measure_score[3] / temp_acc_score  # RSVM  acc#存入本次训练精确度标准差accuracy_std[0].append(accuracy_measure_score[0] * temp_acc_score)  # Knn   stdaccuracy_std[1].append(accuracy_measure_score[1] * temp_acc_score)  # DTree stdaccuracy_std[2].append(accuracy_measure_score[2] * temp_acc_score)  # LSVM  stdaccuracy_std[3].append(accuracy_measure_score[3] * temp_acc_score)  # RSVM  std#计算本次训练的f1分数f1_res_score[0] += fl_measure_score[0] / temp_f1_score  # Knn   f1f1_res_score[1] += fl_measure_score[1] / temp_f1_score  # DTree f1f1_res_score[2] += fl_measure_score[2] / temp_f1_score  # LSVM  f1f1_res_score[3] += fl_measure_score[3] / temp_f1_score  # RSVM  f1#存入本次训练f1分数标准差f1_res_std[0].append(fl_measure_score[0] * temp_f1_score)  # Knn   stdf1_res_std[1].append(fl_measure_score[1] * temp_f1_score)  # DTree stdf1_res_std[2].append(fl_measure_score[2] * temp_f1_score)  # LSVM  stdf1_res_std[3].append(fl_measure_score[3] * temp_f1_score)  # RSVM  std#计算本次训练的roc_aoc分数roc_aoc_res_score[0] += roc_measure_score[0] / temp_roc_score  # Knn   rocroc_aoc_res_score[1] += roc_measure_score[1] / temp_roc_score  # DTree rocroc_aoc_res_score[2] += roc_measure_score[2] / temp_roc_score  # LSVM  rocroc_aoc_res_score[3] += roc_measure_score[3] / temp_roc_score  # RSVM  roc#存入本次训练的roc_aoc分数标准差roc_aoc_res_std[0].append(roc_measure_score[0] * temp_roc_score)  # Knn   stdroc_aoc_res_std[1].append(roc_measure_score[1] * temp_roc_score)  # DTree stdroc_aoc_res_std[2].append(roc_measure_score[2] * temp_roc_score)  # LSVM  stdroc_aoc_res_std[3].append(roc_measure_score[3] * temp_roc_score)  # RSVM  stdknn_acc_res = "&$%.1f" % (accuracy_res_score[0] * 10) + "_{\pm%.2f" % (np.std(accuracy_std[0])) + "}$"knn_f1_res = "&$%.1f" % (f1_res_score[0] * 10) + "_{\pm%.2f" % (np.std(f1_res_std[0])) + "}$"knn_roc_res = "&$%.1f" % (roc_aoc_res_score[0] * 10) + "_{\pm%.2f" % (np.std(roc_aoc_res_std[0])) + "}$"DTree_acc_res = "&$%.1f" % (accuracy_res_score[1] * 10) + "_{\pm%.2f" % (np.std(accuracy_std[1])) + "}$"DTree_f1_res = "&$%.1f" % (f1_res_score[1] * 10) + "_{\pm%.2f" % (np.std(f1_res_std[1])) + "}$"DTree_roc_res = "&$%.1f" % (roc_aoc_res_score[1] * 10) + "_{\pm%.2f" % (np.std(roc_aoc_res_std[1])) + "}$"LSVM_acc_res = "&$%.1f" % (accuracy_res_score[2] * 10) + "_{\pm%.2f" % (np.std(accuracy_std[2])) + "}$"LSVM_f1_res = "&$%.1f" % (f1_res_score[2] * 10) + "_{\pm%.2f" % (np.std(f1_res_std[2])) + "}$"LSVM_roc_res = "&$%.1f" % (roc_aoc_res_score[2] * 10) + "_{\pm%.2f" % (np.std(roc_aoc_res_std[2])) + "}$"RSVM_acc_res = "&$%.1f" % (accuracy_res_score[3] * 10) + "_{\pm%.2f" % (np.std(accuracy_std[3])) + "}$"RSVM_f1_res = "&$%.1f" % (f1_res_score[3] * 10) + "_{\pm%.2f" % (np.std(f1_res_std[3])) + "}$"RSVM_roc_res = "&$%.1f" % (roc_aoc_res_score[3] * 10) + "_{\pm%.2f" % (np.std(roc_aoc_res_std[3])) + "}$"#输出结果print('\t\t\t\t\t', knn_acc_res, '\t', knn_f1_res, '\t', knn_roc_res, '\t',DTree_acc_res, '\t', DTree_f1_res, '\t', DTree_roc_res, '\t',LSVM_acc_res, '\t', LSVM_f1_res, '\t', LSVM_roc_res, '\t',RSVM_acc_res, '\t', RSVM_f1_res, '\t', RSVM_roc_res)if __name__ == '__main__':file_name = "mutagenesis1.mat"para_name = file_name[30:]DIEPredict(file_name,para_name)