Perform Naive-Bayes classification(fitcnb) with non-zero off-diagonal covariance matrix

Greetings,

I use a Bayesian classification model to generate class-conditional probability density functions (PDFs) from a Monte Carlo (MC) simulation (see Fig 1). The different classes have inter-variable correlations such that the covariance matrix has non-zeros on the off-diagonal elements. However, the Bayesian classification model seems to assume that the off-diagonal elements are zero, such that the PDFs for each class are not shaped according to the MC simulated data (see Fig 2); this makes the PDFs look like ellipsoids that are horizontally aligned.

So, how can I specify the covariance elements in the Bayesian classification model when I for instance want to use it to predict a new data set?

Thanks,

Kenneth

Fig 1:

Fig2:

Answer 1

[the cyclist]

Disclaimer: I am not an expert on these methods.

Doesn't the "naive" in naive Bayes specifically mean that the model features are independent from each other (i.e. uncorrelated)? You might need a more sophisticated model.

[simplified]

You are completely right, thank you for reminding me of this essential property!

Answer 2

[Ilya]

To estimate covariance per class, use fitcdiscr with discriminant type 'quadratic'.