For example, the confusion matrix (confMat) is: confMat(1,1) = 104, confMat(1,2) = 1242, confMat(2,1) = 14 and confMat(2,2) = 14199. In matalab using accuracy = mean(diag(confMat)) provides 0.5381. However, using the relation: accuracy=(TP+TN)/(TP+TN+FP+FN)=91.92. I have two classes and unbalanced dataset.
I believe the matrix of confusion in matlab is: confMat(1,1) = TN, confMat (1,2) = FP, confMat (2,1) = FN and confMat(2,2) = TP. Whereas: Horizontal = predicted class and Vertical = real class