Alex,
I gave it a second thought and you are rite in pointing out that I should not be forcing decision boundaries on the posterior predictions.
I have now reduced the number of predictors to two and the accuracy has improved from mid 50’s to high 70’s. However, one of the predictors is continuous and one is discrete ( six distinct values - 1, 2, 3, 4, 5, 6). Is there a different approach in terms of the selection of priors or sampling algorithm when the predictors are comprised of discrete and continuous values?
In addition to the multivariate multi-nomial regression, I have recast the problem to a multivariate-binomial regression. I want to visualize the occurrence probability as a function of the two predictors. Can you do surface plots of posterior predictions using arviz in PyMC3. If so, can you point me to snippet of the code?
I again, want to thank you for all your help and advise.
Mahesh