Educational data mining is a new research area and is used to predict student performance and provides insight that allows educators to plan accordingly. Its results now play an important role in improving educational standards. Specific algorithms for �Features Selection� optimize the classification accuracy of a prediction model. This work introduces a new method based on sparse representation for features selection and reduction that assesses predictive model's accuracy, precision and recall. Different existing features selection methods are fused and passed to a classifier to measure performance using educational datasets. Experimental results are compared to existent features selection techniques and demonstrate that the proposed approach provides superior solution for data fusion and individual (single) predictive outcomes. © 2005 � ongoing JATIT & LLS.