Bayesian network classification using spline-approximated kernel density estimation

The likelihood for patterns of continuous features needed for probabilistic inference in a Bayesian network classifier (BNC) may be computed by kernel density estimation (KDE), letting every pattern influence the shape of the probability density. Although usually leading to accurate estimation, the KDE suffers from computational cost making it unpractical in many real-world applications. We smooth the density using a spline thus requiring for the estimation only very few coefficients rather than the whole training set allowing rapid implementation of the BNC without sacrificing classifier accuracy. Experiments conducted over a several real-world databases reveal acceleration in computational speed, sometimes in several orders of magnitude, in favor of our method making the application of KDE to BNCs practical.