相关向量机 (RVM) -Matlab

叁歲伎倆 2022-09-09 10:52 519阅读 0赞

![cba8e5353f06e92a7ab3679882775faa.png][]

## 主要特点 ##

*  用于二元分类 (RVC) 或回归 (RVR) 的 RVM 模型
 *  多种核函数（线性、高斯、多项式、sigmoid、laplacian）
 *  混合核函数（K =w1×K1+w2×K2+...+wn×Kn）
 *  使用贝叶斯优化、遗传算法和粒子群优化进行参数优化

*  此版本的代码与低于 R2016b 的版本不兼容。
 *  详细应用请看演示。
 *  此代码仅供参考。

使用举例

### 01. 使用 RVM (RVC) 进行分类 ###

使用 RVM 进行分类的演示

clc
    clear all
    close all
    addpath(genpath(pwd))
    
    % use fisheriris dataset
    load fisheriris
    inds = ~strcmp(species, 'setosa');
    data_ = meas(inds, 3:4);
    label_ = species(inds);
    cvIndices = crossvalind('HoldOut', length(data_), 0.3);
    trainData = data_(cvIndices, :);
    trainLabel = label_(cvIndices, :);
    testData = data_(~cvIndices, :);
    testLabel = label_(~cvIndices, :);
    
    % kernel function
    kernel = Kernel('type', 'gaussian', 'gamma', 0.2);
    
    % parameter
    parameter = struct( 'display', 'on',...
                        'type', 'RVC',...
                        'kernelFunc', kernel);
    rvm = BaseRVM(parameter);
    
    % RVM model training, testing, and visualization
    rvm.train(trainData, trainLabel);
    results = rvm.test(testData, testLabel);
    rvm.draw(results)

运行结果：

*** RVM model (classification) train finished ***
    running time            = 0.1604 seconds
    iterations              = 20 
    number of samples       = 70 
    number of RVs           = 2 
    ratio of RVs            = 2.8571% 
    accuracy                = 94.2857%
    
    
    *** RVM model (classification) test finished ***
    running time            = 0.0197 seconds
    number of samples       = 30 
    accuracy                = 96.6667%

### 02. 使用 RVM (RVR) 进行回归 ###

使用 RVM 进行回归的演示

clc
    clear all
    close all
    addpath(genpath(pwd))
    
    % sinc funciton
    load sinc_data
    trainData = x;
    trainLabel = y;
    testData = xt;
    testLabel = yt;
    
    % kernel function
    kernel = Kernel('type', 'gaussian', 'gamma', 0.1);
    
    % parameter
    parameter = struct( 'display', 'on',...
                        'type', 'RVR',...
                        'kernelFunc', kernel);
    rvm = BaseRVM(parameter);
    
    % RVM model training, testing, and visualization
    rvm.train(trainData, trainLabel);
    results = rvm.test(testData, testLabel);
    rvm.draw(results)

运行结果：

*** RVM model (regression) train finished ***
    running time            = 0.1757 seconds
    iterations              = 76 
    number of samples       = 100 
    number of RVs           = 6 
    ratio of RVs            = 6.0000% 
    RMSE                    = 0.1260
    R2                      = 0.8821
    MAE                     = 0.0999
    
    
    *** RVM model (regression) test finished ***
    running time            = 0.0026 seconds
    number of samples       = 50 
    RMSE                    = 0.1424
    R2                      = 0.8553
    MAE                     = 0.1106

### 03. 内核函数 ###

定义了一个名为***Kernel***的类来计算核函数矩阵。

%{
            type   -
            
            linear      :  k(x,y) = x'*y
            polynomial  :  k(x,y) = (γ*x'*y+c)^d
            gaussian    :  k(x,y) = exp(-γ*||x-y||^2)
            sigmoid     :  k(x,y) = tanh(γ*x'*y+c)
            laplacian   :  k(x,y) = exp(-γ*||x-y||)
        
        
            degree -  d
            offset -  c
            gamma  -  γ
    %}
    kernel = Kernel('type', 'gaussian', 'gamma', value);
    kernel = Kernel('type', 'polynomial', 'degree', value);
    kernel = Kernel('type', 'linear');
    kernel = Kernel('type', 'sigmoid', 'gamma', value);
    kernel = Kernel('type', 'laplacian', 'gamma', value);

例如，计算**X**和**Y**之间的核矩阵

X = rand(5, 2);
    Y = rand(3, 2);
    kernel = Kernel('type', 'gaussian', 'gamma', 2);
    kernelMatrix = kernel.computeMatrix(X, Y);
    >> kernelMatrix
    
    kernelMatrix =
    
        0.5684    0.5607    0.4007
        0.4651    0.8383    0.5091
        0.8392    0.7116    0.9834
        0.4731    0.8816    0.8052
        0.5034    0.9807    0.7274

### 04. 混合内核 ###

使用带有 hybrid\_kernel 的 RVM 进行回归的演示 (K = w1 *K1+w2* K2+...+wn\*Kn)

clc
    clear all
    close all
    addpath(genpath(pwd))
    
    % sinc funciton
    load sinc_data
    trainData = x;
    trainLabel = y;
    testData = xt;
    testLabel = yt;
    
    % kernel function
    kernel_1 = Kernel('type', 'gaussian', 'gamma', 0.3);
    kernel_2 = Kernel('type', 'polynomial', 'degree', 2);
    kernelWeight = [0.5, 0.5];
    % parameter
    parameter = struct( 'display', 'on',...
                        'type', 'RVR',...
                        'kernelFunc', [kernel_1, kernel_2],...
                        'kernelWeight', kernelWeight);
    rvm = BaseRVM(parameter);
    
    % RVM model training, testing, and visualization
    rvm.train(trainData, trainLabel);
    results = rvm.test(testData, testLabel);
    rvm.draw(results)

### 05. 单核RVM参数优化 ###

带参数优化的 RVM 模型演示

clc
    clear all
    close all
    addpath(genpath(pwd))
    
    % use fisheriris dataset
    load fisheriris
    inds = ~strcmp(species, 'setosa');
    data_ = meas(inds, 3:4);
    label_ = species(inds);
    cvIndices = crossvalind('HoldOut', length(data_), 0.3);
    trainData = data_(cvIndices, :);
    trainLabel = label_(cvIndices, :);
    testData = data_(~cvIndices, :);
    testLabel = label_(~cvIndices, :);
    
    % kernel function
    kernel = Kernel('type', 'gaussian', 'gamma', 5);
    
    % parameter optimization
    opt.method = 'bayes'; % bayes, ga, pso
    opt.display = 'on';
    opt.iteration = 20;
    
    % parameter
    parameter = struct( 'display', 'on',...
                        'type', 'RVC',...
                        'kernelFunc', kernel,...
                        'optimization', opt);
    rvm = BaseRVM(parameter);
    
    % RVM model training, testing, and visualization
    rvm.train(trainData, trainLabel);
    results = rvm.test(trainData, trainLabel);
    rvm.draw(results)

结果：

*** RVM model (classification) train finished ***
    running time            = 13.3356 seconds
    iterations              = 88 
    number of samples       = 70 
    number of RVs           = 4 
    ratio of RVs            = 5.7143% 
    accuracy                = 97.1429%
    Optimized parameter  table
    
        gaussian_gamma
        ______________
    
            7.8261    
    
    *** RVM model (classification) test finished ***
    running time            = 0.0195 seconds
    number of samples       = 70 
    accuracy                = 97.1429%

### 06. Hybrid-kernel-RVM 参数优化 ###

带参数优化的 RVM 模型演示

%{
        A demo for hybrid-kernel RVM model with Parameter Optimization
    %}
    
    
    clc
    clear all
    close all
    addpath(genpath(pwd))
    
    % data
    load UCI_data
    trainData = x;
    trainLabel = y;
    testData = xt;
    testLabel = yt;
    
    % kernel function
    kernel_1 = Kernel('type', 'gaussian', 'gamma', 0.5);
    kernel_2 = Kernel('type', 'polynomial', 'degree', 2);
    
    % parameter optimization
    opt.method = 'bayes'; % bayes, ga, pso
    opt.display = 'on';
    opt.iteration = 30;
    
    % parameter
    parameter = struct( 'display', 'on',...
                        'type', 'RVR',...
                        'kernelFunc', [kernel_1, kernel_2],...
                        'optimization', opt);
    rvm = BaseRVM(parameter);
    
    % RVM model training, testing, and visualization
    rvm.train(trainData, trainLabel);
    results = rvm.test(testData, testLabel);
    rvm.draw(results)

结果：

*** RVM model (regression) train finished ***
    running time            = 24.4042 seconds
    iterations              = 377 
    number of samples       = 264 
    number of RVs           = 22 
    ratio of RVs            = 8.3333% 
    RMSE                    = 0.4864
    R2                      = 0.7719
    MAE                     = 0.3736
    Optimized parameter  1×6 table
    
        gaussian_gamma    polynomial_gamma    polynomial_offset    polynomial_degree    gaussian_weight    polynomial_weight
        ______________    ________________    _________________    _________________    _______________    _________________
    
            22.315             13.595               44.83                  6               0.042058             0.95794     
    
    
    
    
    *** RVM model (regression) test finished ***
    running time            = 0.0008 seconds
    number of samples       = 112 
    RMSE                    = 0.7400
    R2                      = 0.6668
    MAE                     = 0.4867

### 07. 交叉验证 ###

在此代码中，支持两种交叉验证方法：“K-Folds”和“Holdout”。例如，5-Folds 的交叉验证是

parameter = struct( 'display', 'on',...
                        'type', 'RVC',...
                        'kernelFunc', kernel,...
                        'KFold', 5);

例如，比率为0.3的Holdout方法的交叉验证是

parameter = struct( 'display', 'on',...
                        'type', 'RVC',...
                        'kernelFunc', kernel,...
                        'HoldOut', 0.3);

### 08. 其他 ###

%% custom optimization option
    %{      
        opt.method = 'bayes'; % bayes, ga, pso
        opt.display = 'on';
        opt.iteration = 20;
        opt.point = 10;
    
        % gaussian kernel function
        opt.gaussian.parameterName = {'gamma'};
        opt.gaussian.parameterType = {'real'};
        opt.gaussian.lowerBound = 2^-6;
        opt.gaussian.upperBound = 2^6;
    
        % laplacian kernel function
        opt.laplacian.parameterName = {'gamma'};
        opt.laplacian.parameterType = {'real'};
        opt.laplacian.lowerBound = 2^-6;
        opt.laplacian.upperBound = 2^6;
    
        % polynomial kernel function
        opt.polynomial.parameterName = {'gamma'; 'offset'; 'degree'};
        opt.polynomial.parameterType = {'real'; 'real'; 'integer'};
        opt.polynomial.lowerBound = [2^-6; 2^-6; 1];
        opt.polynomial.upperBound = [2^6; 2^6; 7];
    
        % sigmoid kernel function
        opt.sigmoid.parameterName = {'gamma'; 'offset'};
        opt.sigmoid.parameterType = {'real'; 'real'};
        opt.sigmoid.lowerBound = [2^-6; 2^-6];
        opt.sigmoid.upperBound = [2^6; 2^6];
    %}
    
    %% RVM model parameter
    %{
        'display'    :   'on', 'off'
        'type'       :   'RVR', 'RVC'
        'kernelFunc' :   kernel function
        'KFolds'     :   cross validation, for example, 5
        'HoldOut'    :   cross validation, for example, 0.3
        'freeBasis'  :   'on', 'off'
        'maxIter'    :   max iteration, for example, 1000
    %}

获取代码：[https://ai.52learn.online/code/110][https_ai.52learn.online_code_110]

[cba8e5353f06e92a7ab3679882775faa.png]: /images/20220829/e35edc498fad4d6d9ee5b088b1a5a26b.png
[https_ai.52learn.online_code_110]: https://ai.52learn.online/code/110