二分類問題使用SVM和BP神經網絡的樣例【Speech and not-speech detection】

二分類問題使用SVM和BP神經網絡的樣例【Speech and not-speech detection】 - Carrawayang - 博客園 (cnblogs.com)

數據集采用浙大胡老師的編程作業為例

Assignment 1: Speech and not-speech detection DDL：2017-10-17 Tue. （1）This assignment is carried out by group. You could choose your teammate freely. Each group consists of at most 3 students.

（2）The ‘training.data’ contains the training data. It is from our project to detect whether a person in a video speaks or not. The features are generated in the following way, which may help you making the most of these features. 1、Get the mouth region M from the origin image based on facial landmark detection. 2、Calculate dense optic flow between mouth region of last frame and the current frame and generate a score S that depicts the motion of mouth. 3、Calculate the parameter V which depicts the degree of mouth opening. 4、For frame i, we also calculate the S and V for its previous and next frames. 5、Hence, we generate a 6 dimensional feature vector is X=[Si-1 Si Si+1 Vi-1 Vi Vi+1]. 6、The label is at the end of each line, where +1 represents speaking, and -1 represents not-speaking.

In the training.data, the ratio of positive examples over negative examples is 1:1. Keep this in mind, for if you find your training error or validation error is larger than 50%, that means your solution learns nothing and performs worse than guessing.

（4）You need to write a program to predict speaking or not speaking. For convenience to evaluate your grogram, please use this name for your matlab main function: speakingDetection.m Note about the interface in your function ‘speakingDetection.m’, it should be: function predY= speakingDetection (X) X: The input feature vectors, which is an N6 matrix, where N is the number of feature vectors. predY: The output vector to predict labels of X, which is a N1 vector, and predY(i) = 1 or -1.

Besides MATLAB, you also use Python, as long as you hold the interface protocol above. Note we don’t recommend C/C++.

（5）You can use ANY method to solve this problem.

問題分析

1. 數據解讀：training.data數據為N*7的matrix矩陣，其中6維vector向量為輸入特征 input feature

2. 數據預處理：將training.data讀入，進行dataset的分割，分為6維向量input feature和1維向量label,分割前對數據集進行shuffle，分出測試集以及訓練集

3. 模型選擇：該問題為數據的分類，采用分類算法可以解決，本文以SVM和BP神經網絡為樣例

4. 建立相應模型求解問題

5. 調節參數，達到最優解

BP神經網絡代碼

clc;
clear;
data=importdata('training.data');
P=data(:,1:6);
T=data(:,7);
temp = randperm(size(data,1));
% 訓練集——5000個樣本
P_train = P(temp(1:5000),:)';
T_train = T(temp(1:5000),:)';
P_test = P(temp(end-50:end),:)';
T_test = T(temp(end-50:end),:)';
N=size(T_test,2);
[pn,minp,maxp,tn,mint,maxt]=premnmx(P_train,T_train);
[pn_,minp_,maxp_,tn_,mint_,maxt_]=premnmx(P_test,T_test);
dx=[-1,1;-1,1;-1,1;-1,1;-1,1;-1,1];
net=newff(dx,[6,10,1]);
net.trainParam.goal = 0;
net.trainParam.epochs = 30000;
net.trainParam.lr = 0.03;
net.trainParam.showWindow = 1;
net = train(net,pn,tn);
an = sim(net,pn_);
a=postmnmx(an,mint_,maxt_);
disp(['mse: ' num2str(mse(T_test-an))]);
count=0;
error=0;
for i=1:N
 if abs(a(i)-T_test(i))<0.2
     count=count+1;
 else
     error=error+1;
 end
end
accuracy=count/(count+error)
figure
plot(1:N,T_test,'b*',1:N,a,'ro')
legend('真實值','預測值')
xlabel('預測樣本')
ylabel('實值')

運行結果 accuracy：0.7059 mse: 1.0783

SVM代碼[采用LIBSVM]

data=importdata('training.data');
features=data(:,1:6);%特征列表
classlabel=data(:,7);%對應類別
n = randperm(size(features,1));%隨機產生訓練集和測試集
%% 訓練集--70個樣本
train_features=features(n(1:44000),:);
train_label=classlabel(n(1:44000),:);
%% 測試集--30個樣本
test_features=features(n(44000:end),:);
test_label=classlabel(n(44000:end),:);
%% 數據歸一化
[Train_features,PS] = mapminmax(train_features');
Train_features = Train_features'; 
Test_features = mapminmax('apply',test_features',PS); 
Test_features = Test_features';
%% 創建/訓練SVM模型
model = svmtrain(train_label,Train_features,'-h 0');
%% SVM仿真測試
[predict_train_label] = svmpredict(train_label,Train_features,model);
[predict_test_label] = svmpredict(test_label,Test_features,model);
%% 打印准確率
compare_train = (train_label == predict_train_label);
accuracy_train = sum(compare_train)/size(train_label,1)*100; 
fprintf('訓練集准確率：%f\n',accuracy_train)
compare_test = (test_label == predict_test_label);
accuracy_test = sum(compare_test)/size(test_label,1)*100;
fprintf('測試集准確率：%f\n',accuracy_test)

運行結果

.................*
optimization finished, #iter = 17228
nu = 0.658959
obj = -28684.553581, rho = 4.599546
nSV = 29001, nBSV = 28987
Total nSV = 29001
Accuracy = 71.7273% (31560/44000) (classification)
Accuracy = 71.1948% (435/611) (classification)
訓練集准確率：71.727273
測試集准確率：71.194763

結果分析

兩種模型按照題目要求可以達到錯誤率低於50%的要求，相對而言，SVM在該問題上無論是性能還是效果都略高於BP神經網絡算法，SVM更適用於小樣本的分類問題

 

文件下載：training.data

https://files.cnblogs.com/files/Carraway-Space/training.zip