Current MLL studies show dataset characterization has effect on the performance of certain MLL algorithm. With MLL dataset characteristics as imbalance and with label dependency issues, it is the research hypothesis that dataset linearity characteristic affects algorithm performance. The study used Soil Test Report as the nonlinear dataset for Ensemble of Pruned set. EPS with different base classifier was modified to test the hypothesis. EPS with non-linear base classifier works better in MLL dataset with non-linear in character.