使用Libsvm的平衡精度进行Matlab参数选择

Question

我正在尝试合并这两位代码。 我试图使用平衡精度（BAC）而不是精度来进行参数选择。 我已经在libsvm上下载了补充文件以处理平衡精度。

我用于参数选择的代码是：

prompt ='CROSS VALIDATION MAXIMUM STEP RANGE) ? ';
maxstep = input (prompt);
stepSize = 1;
log2c_list = -maxstep:stepSize:maxstep;
log2g_list = -maxstep:stepSize:maxstep;
maxc = max(log2c_list);
maxg = max(log2c_list);
numLog2c = length(log2c_list);
numLog2g = length(log2g_list);
cvMatrix = zeros(numLog2c,numLog2g);
bestcv = 0;
for i = 1:numLog2c
    log2c = log2c_list(i);
    for j = 1:numLog2g
        log2g = log2g_list(j);
        % -v 3 --> 3-fold cross validation
        param = ['-q -v 3 -c ', num2str(2^log2c), ' -g ', num2str(2^log2g)];
        cv = svmtrain(class_vector_train, predictors_matrix_train, param);
        cvMatrix(i,j) = cv;
        if (cv >= bestcv),
            bestcv = cv; bestLog2c = log2c; bestLog2g = log2g;
        end
        % fprintf('%g %g %g (best c=%g, g=%g, rate=%g)\n', log2c, log2g, cv, bestc, bestg, bestcv);
    end
end
disp(['CV scale1: best log2c:',num2str(bestLog2c),' best log2g:',num2str(bestLog2g),' accuracy:',num2str(bestcv),'%']);

为了获得我使用的天平精度：

do_binary_cross_validation(class_vector_train, predictors_matrix_train,'-c 1 -g 2',5);
model = svmtrain(class_vector_train, predictors_matrix_train);
[predicted_class_test, evaluation_results, decision_values] = do_binary_predict(class_vector_test, predictors_matrix_test, model);

但是我找不到平衡的精度度量。

Answer 1

使用功能bac您的validate_function.m文件：

function ret = bac(dec, label)
tp = sum(label == 1 & dec >= 0);
tn = sum(label == -1 & dec < 0);
tp_fn = sum(label == 1);
tn_fp = sum(label == -1);
if tp_fn == 0;
  disp(sprintf('warning: No positive true label.'));
  sensitivity = 0;
else
  sensitivity = tp / tp_fn;
end
if tn_fp == 0;
  disp(sprintf('warning: No negative true label.'));
  specificity = 0;
else
  specificity = tn / tn_fp;
end
ret = (sensitivity + specificity) / 2;
disp(sprintf('BAC = %g', ret));

这将计算BAC (Balanced ACcuracy) = (Sensitivity + Specificity) / 2 ，其中Sensitivity = true_positive / (true_positive + false_negative)和Specificity = true_negative / (true_negative + false_positive)