newton
05-22-2010, 07:26 AM
I am getting positive loglikelihood sometimes (with gaumixmod.h). This is happening when the gaussian density is coming out to be more than one. I am not being able to decide what to do with the positive log-likelihood model solutions. Shall I accept them or discard them?
I noticed that sometimes a particular component in the Gaussian mixtures attract very few datapoints and the sparsity leads to Gaussian density being greater than one. But this is always not the case, implying that there may be other factors driving the log-likelihood being positive.
Can anybody suggest what to do with positive log-likelihood solutions? Do they really mean better solutions as compared to the negative log-likelihood ones or do they happen because of some rare abnormality and hence need rejection?
-Newton
I noticed that sometimes a particular component in the Gaussian mixtures attract very few datapoints and the sparsity leads to Gaussian density being greater than one. But this is always not the case, implying that there may be other factors driving the log-likelihood being positive.
Can anybody suggest what to do with positive log-likelihood solutions? Do they really mean better solutions as compared to the negative log-likelihood ones or do they happen because of some rare abnormality and hence need rejection?
-Newton