In density estimation, it is just another Santa Claus hat on top of each observation, whose bandwith you hopefully optimize e.g. by 10-fold CV, or preferably 10 times 10-fold CV. See original paper (cited by 977): John&Langley: Estimating continuous distributions in Bayesian classifiers. In Proceedings of the Eleventh Confrence on Uncertainty in Artifial Intelligence, Morgann Kaufmann Publishers, San Mateo, 1995.
The whole algorithm is found e.g in RapidMiner freeware. Cheers P