根据每个输入元素返回 tp, tn, fn, fp

Question

I have a csv file with true and predicted labels (4 classes) associated with an ID. The csv file looks like this:

task_id,labels_true,labels_pred
76017-126511-18,2,2
76017-126512-18,0,3
76017-126513-18,2,2
76018-126511-18,2,2
76018-126512-18,2,2
76018-126513-18,2,1
76019-126511-18,2,2
76019-126512-18,1,0

I am using the confusion matrix from sklearn.metrics

y_true = df["labels_true"]
y_pred = df["labels_pred"]

cnf_matrix = confusion_matrix(y_true, y_pred, labels=[0,1,2,3])

It returns an array as follows:

[[ 554    1   28    0]
[  15  1375   43    0]
[  42   476 2263    0]
[   0    0    0    0]]

My aim is to return a list with each element ID associated with the respective tp, tn, fp, fn values like this:

task_id,labels_true,labels_pred, cm
76017-126511-18,2,2, tp 
76017-126513-18,2,2, tp
76018-126511-18,2,2, tp

Answer 1

It's a multi class confusion matrix. True/False positives are used for binary classification problems. What you can do is to ecncode your labels as a binary values for example (classes 1,2,3 encoded as 1) and recalculate the confusion matrix.

Answer 2

TL;DR : For multi-class cases, this is not possible.

---

As already suggested, the very notions of True Positives (TP), True Negatives (TN), False Positives (FP), and False Negatives (FN) come from binary classification settings; they can indeed be used in multi-class classification, as shown here , but in such cases the notions are not a straightforward extension of the binary case, making what you ask here actually impossible.

In multi-class classification, all these notions are defined and calculated per class . And this renders any effort to uniquely identify a sample as being in one and only one of these categories (TP, FP, TN, FN) impossible.

Let's demonstrate this with some examples, using your case (4 classes [0, 1, 2, 3] ).

Take a misclassified sample first, eg:

True label:      0
Predicted label: 3

• From the point of view (POV) of class 0 , this is a False Negative (FN): prediction is not 0 , as it should be
• From the POV of class 1 , this is a True Negative: it is not 1 , and it has correctly been classified as not 1
• From the POV of class 2 , this is again a True Negative (TN): it is not 2 , and it has correctly been classified as not 2
• From the POV of class 3 , this is a False Positive (FP): it has been wrongly classified as 3 without being so

Similar is the case for a correct classification, say

True label:      2
Predicted label: 2

• From the POV of class 0 , this is a True Negative (TN): it is not 0 , and it has correctly been classified as not 0
• From the POV of class 1 , this is a True Negative (TN): it is not 1 , and it has correctly been classified as not 1
• From the POV of class 2 , this is a True Positive (TP)
• From the POV of class 3 , this is a True Negative (TN): it is not 3 , and it has correctly been classified as not 3

---

Given this exposition, it should hopefully be clear that what you ask is actually not possible in the multi-class case.