How to get confusion matrix with 100% precision or 100% recall in Weka Here is the original confusion matrix:  
    | Yes |  No  |             
 -----------------        The top right corner:   False Negative
 Yes|1000 | 200  |        The top left corner:    True Positive
 -----------------        The bottom right corner:True Negative
  No|200  | 500  |        The bottom left corner: False Positive

I expect to get these :


*

*Basically, I am looking for an option or command in the classifier that can enforce WEKA not misclassifying the 200 TP instances to the FN section. In other words, the number of TP instances will be 1000 + 200 = 1200, and number of FN instances will be 0 or 100% precision.
   |  Yes  |  No  |     
-------------------     
Yes| 1200  |  0   |     
-------------------     
No | 200   | 500  |     


*Same idea for the second part, all the FP instances are moved to TN section or 100% recall.
   |  Yes  |  No  |    
-------------------    
Yes| 1000  | 200  |
-------------------
No |   0   | 500  |

Is there any way to implement this idea in WEKA?
 A: Have you tried playing with the Cost Matrix (Cost-sensitive evaluation)? It can not guarantee a priori your outcome, but can force Weka to be more sensitive to a particular type or error.
A: Intuitively, this would require that the classifier somehow “knows” the true values (to be able to separate false from true negatives and reclassify the former as positive instances). But if that were the case, the whole classification exercise would seem quite unnecessary. In effect, you seem to be asking “can I create a perfect classifier to correct my output after the fact?”, which is obviously impossible, so what you are trying to achieve is unclear to me.
What you can obviously do though is this:
    | Yes |  No  |             
 -----------------
 Yes|1200 |   0  |
 -----------------
  No|700  |   0  |

If a classifier classifies everything as “positive”, recall is automatically 100% (precision is likely to be bad but really depend on the proportion of positive exemplars in your test data set). It's possible to do something similar for specificity (by classifying everything as “negative”) but not for precision.
