I guess so, since you have 4 unknowns and 4 equations.

Try using standard definitions and solve for entries in the confusion matrix.

TPR = Am_TP/(Am_TP + Am_FN), where Am is Amount

TNR = Am_TN/(Am_TN + Am_FP)

Accuracy = (Am_TP + Am_TN)/(Am_TP + Am_TN +Am_FP + Am_FN)

So, generally speaking, it is impossible. But if you know (Am_TP + Am_FN) - amount of all positive objects or (Am_TN + Am_FP) - amount of all negative objects, or some else (e.g., Am_TP/Am_TN), it is, of course, possible.

Not sure if you can derive accuracy (ACC) from TPR and TNR, but if you know the total number of Positive and the total number Negative examples in your test data (Lets call it A and B, respectively), you may be able to compute ACC. Let's call the number of true positives given by your algorithm a and the number of true negatives b.

Since:

TPR = a/A

and

TNR = b/B

Then you can infer a and b and them compute ACC:

ACC = (a + b)/(A + B)

Hope it can help, but I think you also should be able to set Weka to output ACC, which would make you life easier.

Regards,

Ricardo

No, you could not. You should use additional data.