libsvm多標籤分類輸出預測概率

Question

我正在嘗試使用libsvm（與Matlab接口）來運行一些多標籤分類問題。下面是一個使用IRIS數據的一些玩具問題：

load fisheriris; 

featuresTraining      = [meas(1:30,:); meas(51:80,:); meas(101:130,:)]; 
featureSelectedTraining     = featuresTraining(:,1:3); 

groundTruthGroupTraining    = [species(1:30,:); species(51:80,:); species(101:130,:)]; 
[~, ~, groundTruthGroupNumTraining]  = unique(groundTruthGroupTraining); 

featuresTesting       = [meas(31:50,:); meas(81:100,:); meas(131:150,:)]; 
featureSelectedTesting     = featuresTesting(:,1:3); 

groundTruthGroupTesting     = [species(31:50,:); species(81:100,:); species(131:150,:)]; 
[~, ~, groundTruthGroupNumTesting]  = unique(groundTruthGroupTesting); 

% Train the classifier 
optsStruct        = ['-c ', num2str(2), ' -g ', num2str(4), '-b ', 1]; 
SVMClassifierObject      = svmtrain(groundTruthGroupNumTraining, featureSelectedTraining, optsStruct); 

optsStruct        = ['-b ', 1]; 
[predLabelTesting, predictAccuracyTesting, ... 
    predictScoresTesting]    = svmpredict(groundTruthGroupNumTesting, featureSelectedTesting, SVMClassifierObject, optsStruct); 

但是，對於預測概率我有（結果的前12行顯示在這裏）

1.08812899093155 1.09025554950852 -0.0140009056912001 
0.948911671379753 0.947899227815959 -0.0140009056926024 
0.521486301840914 0.509673405799383 -0.0140009056926027 
0.914684487894784 0.912534150299246 -0.0140009056926027 
1.17426551505833 1.17855350325579 -0.0140009056925103 
0.567801459258613 0.557077025701113 -0.0140009056926027 
0.506405203427106 0.494342606399178 -0.0140009056926027 
0.930191457490471 0.928343421250020 -0.0140009056926027 
1.16990617214906 1.17412523596840 -0.0140009056926026 
1.16558843984163 1.16986137054312 -0.0140009056926015 
0.879648874624610 0.876614924593740 -0.0140009056926027 
-0.151223818963057 -0.179682730685229 -0.0140009056925999 

我感到困惑的是如何某些概率大於1，其中一些是負數？

然而，預測標籤似乎相當準確：

1 
1 
1 
1 
1 
1 
1 
1 
1 
1 
1 
3 

與

Accuracy = 93.3333% (56/60) (classification) 

最終輸出，那麼如何解釋預測概率的結果嗎？非常感謝。 A.

Answer 1

svm的輸出不是概率！

得分的符號表明它是屬於A類還是B類。如果得分是1或-1，則它在邊界上，儘管這對知道並不特別有用。

如果您確實需要概率，則可以使用Platt scaling進行轉換。你基本上應用一個sigmoid函數。

Answer 2

據我所知，這個答案可能太晚了，但它可能有益於遇到同樣問題的人。

libsvm實際上可以產生概率，使用選項'-b'。

我認爲你犯的錯誤是你定義optsStruct變量的方式。它應該像這樣定義：['-b ' num2str(1)]或['-b 1']。

這同樣適用於發送到svmtrain的選項。