OLAP techniques provide efficient solutions to navigate through data cubes. However, they are not equipped with frameworks that empower user investigation of interesting information. They are restricted to exploration tasks. Recently, various studies have been trying to extend OLAP to new capabilities by coupling it with data mining algorithms. However, most of these algorithms are not designed to deal with sparsity, which is an unavoidable consequence of the multidimensional structure of OLAP cubes. In [1], we proposed a novel approach that embeds Multilayer Perceptrons into OLAP environment to extend it to prediction. This approach has largely met its goals with limited sparsity cubes. However, its performances have decreased progressively with the increase of cube sparsity. In this paper, we propose a substantially modified version of our previous approach called NAP-SC (Neural Approach for Prediction over Sparse Cubes). Its main contribution consists in minimizing sparsity effect on measures prediction process through the application of a cube transformation step, based on a dedicated aggregation technique. Carried out experiments demonstrate the effectiveness and the robustness of NAP-SC against high sparsity data cubes