Summary

This chapter teach us about systematically improving your learning algorithm. The videos for this week will teach you how to tell when a learning algorithm is doing poorly, and describe the ‘best practices’ for how to ‘debug’ your learning algorithm and go about improving its performance.

0.

第七周作业的题目：点击查看

1. Support Vector Machines

1.1 Training Linear SVM

C = 1;
//C = 100;
model = svmTrain(X, y, C, @linearKernel, 1e-3, 20);
visualizeBoundaryLinear(X, y, model);

1.2 SVM with Gaussian Kernels

完成文件“gaussianKernel.m”计算两个样本之间的Gaussian kernel高斯核函数,即两个点的相似度。

//gaussianKernel.m

sim = exp(-sum((x1-x2).^2)/(2*sigma^2));

Training SVM with RBF Kernel （dataset2）

model= svmTrain(X, y, C, @(x1, x2) gaussianKernel(x1, x2, sigma)); 
visualizeBoundary(X, y, model);

Training SVM with RBF Kernel （dataset3）

尝试不同的参数C和sigma，找出较优参数

添加代码至dataset3Params.m找出最优参数

//dataset3Params.m

param = [0.01 0.03 0.1 0.3 1 3 10 30]';

C = 0.01;
sigma = 0.01;
model= svmTrain(X, y, C, @(x1, x2) gaussianKernel(x1, x2, sigma)); 
predictions = svmPredict(model, Xval);  
min_error = mean(double(predictions ~= yval));

for i = 1:length(param)
   for j = 1:length(param)
          C = param(i);
          sigma = param(j);
       model= svmTrain(X, y, C, @(x1, x2) gaussianKernel(x1, x2, sigma)); 
       predictions = svmPredict(model,Xval);
       prediction_error = mean(double(predictions ~=yval));
       if (prediction_error < min_error)
               cm = C;
               sigmam=sigma;
               min_error = mean(double(predictions ~= yval));
           end

    end
end

C = cm;
sigma = sigmam;

运行结果

2 Spam Classification

spam (y = 1) or non-spam (y = 0)

2.1 Preprocessing Emails

完成processEmail.m

//processEmail.m

for i = 1:length(vocabList)
    if strcmp(str,vocabList{i})
        word_indices = [word_indices;i];
    end
end

运行结果

2.2 Extracting Features from Emails

完成emailFeatures.m

//emailFeatures.m
x(word_indices) = 1;

运行结果

2.3 Training SVM for Spam Classification

可以训练一个线性分类器，来决定某个邮件是否是垃圾邮件。

2.4 Top Predictors for Spam

根据前文训练出来的model，获得vocabulary list的权重并排序

运行结果