The development of a fully automatic facial expression recognition system is an open problem. Its implications are very important, with applications ranging from machine intelligence and interaction to psychology research. In order to obtain a viable system, it is necessary to get valid parameters to characterize the facial expression in an image or a video sequence. Several different techniques have been implemented, using global-based, local-based and hybrid methods. In our work we developed a new algorithm based on POEM algorithms. We tested the performance using the Cohn-Kanade database and we compared the results with algorithms using geometric features and regular LBP patterns. Additionally, since the parameters have high linear and non-linear dependence they don't have an homogeneous statistic importance as descriptors, so we performed data mining processing. Our results show that POEM-based algorithms have high performance and low cost, even with low resolution images, outperforming most of traditional state of the art works. Preliminary tests also show the viability of using meta classifiers in order to further improve the performance.