Performance of decision tree algorithm with randomly distributed feature space

This study investigates the performance of well-known tree-based algorithms, namely the J48 and Random Forest (RF), on data sets with various feature space distributions. The feature space was simulated using logistic regression due to its simplicity in linearly separable scenarios mainly to reflect a bell-shaped (Normal), right (Gamma) and left (Beta) skewness of real data sets. Both algorithms were compared through accuracy. Simulation results indicate that with normally distributed feature space of balanced class data sets (Sim1, Sim2), J48 performed less accurately than RF. Meanwhile, both algorithms were found to be equally accurate in rather randomly distributed feature space with imbalanced class. This is expected since most algorithms were designed to detect the majority groups for high classification rate. Nevertheless, the simulation procedure highlights the importance of investigating the feature space structure in model performance particularly in unstructured or noisy environments due to reliance on feature space conditions. Future works would be to consider the overall structure of the data sets including class distributions and number of splits as decision trees rely heavily on the branches for classification tasks.