1
$\begingroup$

I am trying to combine the predicted class labels of 3 different classifiers (for example SVM, Naive Bayes etc ) so that it will eliminate the weakness of each individual classifier. I am trying to use dempster Shafer evidence theory in order to fuse this predicted class labels in Matlab. Unfortunately, I could not find a single implementation of DSET in Matlab. So, can anyone please tell me how to use DSET to fuse class labels of different classifiers in Matlab?

$\endgroup$
2
$\begingroup$

If your really understand the Dempster Shafer Evidence Theory you can do it by yourself, and it is just couple of lines of code. I think this may be the reason that you cannot find it in a package, because it is trivial.

On the other hand, I believe Dempster Shafer Evidence Theory is better for "belief aggregation". For example, the widely used example in Dempster Shafer Evidence Theory is doctor dignosis, which doctor believe 60% the patient has disease A and 40% disease B. Note, there is no "pior" data here, just numbers from human experts.

In your case, it may be better to use Bayesian approch instead of Dempster Shafer Evidence Theory. In addition search for "bagging" or "ensemble leanring" to see how others to this task.

$\endgroup$
  • $\begingroup$ What about weighted majority voting? is that a good choice? Bagging and Boosting are improving my predictions but my I am asked to try out DSET as well. $\endgroup$ – Ambarish Jan 10 '18 at 20:03
  • 1
    $\begingroup$ @Ambarish when I was working on a research project, the mind set of "try everything and see which is the best" is not a good one. Very hard to defense. When we try something it is better to have a good reason. Back to weighted average/voting. How to decide on weights? ad-hocly assign? or you want to train the weights also in some way. $\endgroup$ – hxd1011 Jan 10 '18 at 20:43
  • $\begingroup$ That makes sense!! As for weighted voting, I am planning to assign more weight to my best classifier and less to the poor performing one and so on. You can say it is more of an ad-hoc assigning. $\endgroup$ – Ambarish Jan 10 '18 at 21:07
  • 1
    $\begingroup$ @Ambarish yes, you are free to design whatever rules for the system and empirically evaluate it to see if it works in your data. But it is hard to justify if you want to publish a paper, if there is no a good reason for the design. (when I say good reason, performing well on your data set is not a good reason.) Think about you go gambling and won a lot money, it cannot justify you have a good strategy, but pure luck in one shot. $\endgroup$ – hxd1011 Jan 10 '18 at 21:19

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.