In this paper, we propose a statistically flexible feature mapping technique for count data which are very common in data analysis and pattern recognition applications. In particular, we are interested in supervised learning to improve the existing non-linear classification techniques using Support Vector Machine (SVM). Perfect representation of the data through a kernel function is strongly dependent on the structure of the data. Thus, choosing an appropriate kernel function or feature mapping technique is necessary to get higher accuracy and in this paper, we address this problem by proposing a new feature mapping function derived from Dirichlet Multinomial and Generalized Dirichlet Multinomial distributions for count data. In order to determine the parameters of the distributions, two approaches namely Expectation Maximization (EM) and Minorization-Maximization (MM) are employed. We evaluate our model by experimenting on two different classification tasks concerning natural scene recognition from images and human action recognition from videos. The results show that, incorporating proposed feature mapping technique with different kernels comparatively gives better results than the base kernels in different classification tasks.