MATLAB机器学习决策树网格法优化参数学习笔记

适用于2017b版本以后的版本，之前的老版本可能有所出入

打开matlab机器学习工具箱选择决策树模型进行训练

选择准确度较高的决策树进行导出，点击右上角生成函数

导出后将已知变量进行替换(可以参考我ROC那篇博文）

确定要调整的参数，本文以下图箭头所指两参数为例：

 

这两个参数分别为最大分裂数和分裂准则，为了找到具体的分裂准则，我们打开fitctree函数的内部，步骤如下：

ctrl+F进行搜索即可： 

 

 然后将内容放入元胞数组中，并设置最大分裂次数的搜索范围为1到30：

 在代码外层添加两个循环体i，j，然后修改参数：

修改完毕后再循环末尾储存每次的结果，红框内第一行为求混淆矩阵的函数，第二行用于求F1分数，其中F1分数位于stats结构体中microAVG的最后一行：

运行程序后寻找最大的F1分数的位置：

 

如果找的较好的F1分数，可在代码最前边加上rng随机数种子保存起来，否则下次运行会丢失结果。从上图中可以看出最大分裂次数为4、分裂准则为twoing时F1分数最好。

利用heatmap进行搜索可视化：

 利用之前导出的函数在修改参数为最佳参数后进行预测，预测时将下图注释部分进行注释，然后加一行代码：trainedClassifier.predictFcn(data2)  %  data2为需要预测的数据

 

 源代码（其中data1为训练集，data2为需要预测的数据）：

%% 网格法过程
trainingData = data1;
inputTable = trainingData;
predictorNames = {'VarName1', 'VarName2', 'VarName3', 'VarName4'};
predictors = inputTable(:, predictorNames);
response = inputTable.VarName5;
isCategoricalPredictor = [false, false, false, false];

% 下面这两个变量在计算F1分数时会用到
ClassNames = unique(response); % y中各类的名称
group = response;  % y所在的那一列（真实的类别）
% Train a classifier
% This code specifies all the classifier options and trains the classifier.

SplitCriterion = {'gdi','twoing','deviance'}; %分裂准则有三个可以选的，这里用元胞数组保存
MaxNumSplits = 1:30;  % 最大分裂数（因为这个问题比较简单，我这里搜索设置的最大分裂数的上界是30）
num_i = length(SplitCriterion);  % 第一个超参数SplitCriterion的可能性有3种
num_j = length(MaxNumSplits);  % 第二个超参数MaxNumSplits的可能性有30种
MICRO_F1_SCORE = zeros(num_i,num_j); % 初始化最后得到的结果（初始化是为了加快代码运行速度）
mywaitbar = waitbar(0);   % 设置一个进度条
TOTAL_NUM = num_i*num_j;  % 总共要计算多少次
now_num = 0; % 已经计算了多少次
for i = 1:num_i
    for j = 1:num_j
        rng(520)  % 设定随机数种子，保证结果的可重复性(这里的520也可以换成其他的数字)
        %   注意，如果超参数是字符类型的话，它是被保存在元胞数组中的，取出时需要使用大括号才能得到字符型
        classificationTree = fitctree(...
            predictors, ...
            response, ...
            'SplitCriterion', SplitCriterion{i}, ...
            'MaxNumSplits', MaxNumSplits(j), ...
            'Surrogate', 'off', ...
            'ClassNames', categorical({'变色鸢尾'; '山鸢尾'; '维吉尼亚鸢尾'}));
        
        %       'SplitCriterion' - Criterion for choosing a split. One of 'gdi'
        %                        (Gini's diversity index), 'twoing' for the twoing
        %                        rule, or 'deviance' for maximum deviance reduction
        %                        (also known as cross-entropy). Default: 'gdi'
        
        %       'MaxNumSplits' - Maximal number of decision splits (or branch
        %                        nodes) per tree. Default: size(X,1)-1
        
        % Create the result struct with predict function
        predictorExtractionFcn = @(t) t(:, predictorNames);
        treePredictFcn = @(x) predict(classificationTree, x);
        trainedClassifier.predictFcn = @(x) treePredictFcn(predictorExtractionFcn(x));
        
        % Add additional fields to the result struct
        trainedClassifier.RequiredVariables = {'VarName1', 'VarName2', 'VarName3', 'VarName4'};
        trainedClassifier.ClassificationTree = classificationTree;
        trainedClassifier.About = 'This struct is a trained model exported from Classification Learner R2017a.';
        trainedClassifier.HowToPredict = sprintf('To make predictions on a new table, T, use: \n  yfit = c.predictFcn(T) \nreplacing ''c'' with the name of the variable that is this struct, e.g. ''trainedModel''. \n \nThe table, T, must contain the variables returned by: \n  c.RequiredVariables \nVariable formats (e.g. matrix/vector, datatype) must match the original training data. \nAdditional variables are ignored. \n \nFor more information, see <a href="matlab:helpview(fullfile(docroot, ''stats'', ''stats.map''), ''appclassification_exportmodeltoworkspace'')">How to predict using an exported model</a>.');
        
        % Extract predictors and response
        % This code processes the data into the right shape for training the
        % model.
        inputTable = trainingData;
        predictorNames = {'VarName1', 'VarName2', 'VarName3', 'VarName4'};
        predictors = inputTable(:, predictorNames);
        response = inputTable.VarName5;
        isCategoricalPredictor = [false, false, false, false];
        
        % Perform cross-validation
        partitionedModel = crossval(trainedClassifier.ClassificationTree, 'KFold', 5);
        
        % Compute validation predictions
        [validationPredictions, validationScores] = kfoldPredict(partitionedModel);
        
        % Compute validation accuracy
        validationAccuracy = 1 - kfoldLoss(partitionedModel, 'LossFun', 'ClassifError');
        
        % 下面就是计算每一次循环得到的micro F1分数
        C = confusionmat(group,validationPredictions,'Order',ClassNames);
        stats = statsOfMeasure(C);
        MICRO_F1_SCORE(i,j) = stats.microAVG(end);
        
        % 更新进度条
        now_num = now_num+1;
        mystr=['计算中...',num2str(100*now_num/TOTAL_NUM),'%'];
        waitbar(now_num/TOTAL_NUM,mywaitbar,mystr);
        
    end
end









%% 热力图

MICRO_F1_SCORE
% figure(5)  % 画一个热力图
% heatmap(MaxNumSplits,SplitCriterion,MICRO_F1_SCORE)
figure(6) % 第一列和其他相比太小了 去掉第一列重画
h_graph = heatmap(MaxNumSplits(2:end),SplitCriterion,MICRO_F1_SCORE(:,2:end));
h_graph.XLabel = '最大分裂数';
h_graph.YLabel = '分裂准则';








%% 保存最佳位置
best_micro_f1_score = max(MICRO_F1_SCORE(:))
%找到MICRO_F1_SCORE第一次出现最大值的位置
[r,c] = find(MICRO_F1_SCORE == best_micro_f1_score,1); 
best_SplitCriterion = SplitCriterion{r} % 分裂准则，元胞数组用大括号取出元素
best_MaxNumSplits = MaxNumSplits(c)  % 决策树的最大分裂数









%%下面我们使用最优模型进行预测：
trainingData = data1;
inputTable = trainingData;
predictorNames = {'VarName1', 'VarName2', 'VarName3', 'VarName4'};
predictors = inputTable(:, predictorNames);
response = inputTable.VarName5;
isCategoricalPredictor = [false, false, false, false];

% Train a classifier
% This code specifies all the classifier options and trains the classifier.
classificationTree = fitctree(...
    predictors, ...
    response, ...
    'SplitCriterion', best_SplitCriterion, ...
    'MaxNumSplits', best_MaxNumSplits, ...
    'Surrogate', 'off', ...
    'ClassNames', categorical({'变色鸢尾'; '山鸢尾'; '维吉尼亚鸢尾'}));

% Create the result struct with predict function
predictorExtractionFcn = @(t) t(:, predictorNames);
treePredictFcn = @(x) predict(classificationTree, x);
trainedClassifier.predictFcn = @(x) treePredictFcn(predictorExtractionFcn(x));

% Add additional fields to the result struct
% trainedClassifier.RequiredVariables = {'VarName1', 'VarName2', 'VarName3', 'VarName4'};
% trainedClassifier.ClassificationTree = classificationTree;
% trainedClassifier.About = 'This struct is a trained model exported from Classification Learner R2017a.';
% trainedClassifier.HowToPredict = sprintf('To make predictions on a new table, T, use: \n  yfit = c.predictFcn(T) \nreplacing ''c'' with the name of the variable that is this struct, e.g. ''trainedModel''. \n \nThe table, T, must contain the variables returned by: \n  c.RequiredVariables \nVariable formats (e.g. matrix/vector, datatype) must match the original training data. \nAdditional variables are ignored. \n \nFor more information, see <a href="matlab:helpview(fullfile(docroot, ''stats'', ''stats.map''), ''appclassification_exportmodeltoworkspace'')">How to predict using an exported model</a>.');
% 
% % Extract predictors and response
% % This code processes the data into the right shape for training the
% % model.
% inputTable = trainingData;
% predictorNames = {'VarName1', 'VarName2', 'VarName3', 'VarName4'};
% predictors = inputTable(:, predictorNames);
% response = inputTable.VarName5;
% isCategoricalPredictor = [false, false, false, false];
% 
% % Perform cross-validation
% partitionedModel = crossval(trainedClassifier.ClassificationTree, 'KFold', 5);
% 
% % Compute validation predictions
% [validationPredictions, validationScores] = kfoldPredict(partitionedModel);
% 
% % Compute validation accuracy
% validationAccuracy = 1 - kfoldLoss(partitionedModel, 'LossFun', 'ClassifError');

predict_final_result=trainedClassifier.predictFcn(data2);%预测的结果

 参考：代码参考数学建模清风老师的代码，本文章仅用于整理分享学习笔记，无任何商业用途，若有意见可联系本人删除