In this work, we propose to estimate the density function of heavy tailed data with non-negative support. As the heavy tailed data are characterized by rare observations in the tail, we propose to classify them into two subsets with high and low density, using k-means method, an unsupervised machine learning classification method. To this end, we construct a new estimator that combines two asymmetric BSPE and gamma kernels. However, the smoothing parameter will be estimated by the adaptive Bayesian approach, developed using the two proposed kernels and the classical UCV method. A comparative study between the proposed estimator and the classical estimator on simulated and real data is performed to evaluate their performance.