Comparative study of traditional Bayesian algorithm and MassBayes algorithm using Pendigits dataset

The existing generative classifiers (eg. Naïve Bayes) estimate joint probability distribution p(x,y) or likelihood p(x|y) with the help of different density estimators, which are not suitable for large data sets due to their high time and space complexities. These classifiers also make different assumptions; allow limited dependencies among attributes and estimate one-dimensional likelihood. A new generative classifier known as MassBayes, works without making any assumptions and estimates multi-dimensional likelihood. We evaluate the MassBayes and two different versions of Naïve Bayes (Naïve Bayes using Kernel Density Estimator and Naïve Bayes using Discretisation) for Pendigits data set. Our evaluation shows that MassBayes can work efficiently on large and multi-dimensional datasets. MassBayes gives better classification accuracy than the other existing generative classifiers.