Hmm, what do you mean by combine? Can you explain the setup of the analysis better?

You need to bring more input or at least examples. Usually, you have only one confusion matrix even with several classes otherwise it means that you are not studying the same dataset or population and it does not make sense to combine the studies.

Do you have an ensemble in mind?

@Kasper Christensen..

From 3 confusion matrix I need to give final output in one confusion matrix.

A confusion matrix is a summarized view of the output of a classifier  (predicted class) vs. the real class (gold standard).

If your class space is the same across the different confusion matrices (I.e.  the class labels are exactly the same) then you could just sum the matrices. I'm not sure why would you do that, unless you want a higher level summary of the classifier performance across different  datasets.

I'm not sure this is what you want, but in information retrieval people do micro and macro averaging for computing precision, recall, fmeasure, etc. In micro average you average true/falses positive/negatives and then compute the measures, and in macro average, you compute the measure for each matrix, and then average the results. See

http://rushdishams.blogspot.com.br/2011/08/micro-and-macro-average-of-precision.html

However, if you have a multiclass problem, you can derivate a multiclass confusion matrix. See, eg.

http://www.cs.odu.edu/~mukka/cs495s13/Lecturenotes/Chapter5/recallprecision.pdf

@Francisco Javier Guzman

Thank you Sir for you reply. Yes, I need to give higher level summary of different confusion matrix that works on different dataset having same domain and same number of class classification(positive, negative)

@Ronaldo Cristiano Prati 

Thank You for your reply. And yes it was helpful.

In your previous question you had mentioned that you are using movie review datasets. If you want to use same datasets in confusion matrix for evaluation, then there is no need to merge them, just calculate the average of precision, recall and F-measure for different datasets. My current research is also sentiment analysis and I found researchers adopting the same criteria to compare their results.

If this is not the case which I understand then your problem needs to be clarified.