I am using Matlab Neural Network toolbox for a classification problem. Now considering a single set of data, if the inbuilt neural network is trained and classified with same data multiple number of time, different accuracy and different confusion matrix is obtained. Now which result should I take? Should I take all the vales obtained in all the training instances and average them fix on one particular result?
I can't check at the moment (no Matlab at hand), but I suppose the differences come from the different random seeds used to initialize the neural networks (at least this is the only part which i can think of that has a random component). I would suggest predicting class probabilities, averaging those and then viewing the resulting confusion matrix of the "averaged" prediction. This way you - to a degree - mitigating the effect of randomness resulting from different initializations of the weights.