Rotation Double Random Forest Algorithm to Predict The Food Insecurity Status of Households

Abstract

The ensemble tree method has been proven to handle classification problems well. The strength of the ensemble tree technique lies in the diversity and independence between each tree. Increasing the diversity of mutually independent decision trees improves model performance. Various studies propose the development of ensemble tree-based models by forming algorithms that create decision trees that are formed independently of each other and have various inputs. These include random forest (RF), rotation forest (RoF), double random forest (DRF), and the latest is rotation double random forest (RoDRF). RoDRF rotates or transforms data intending to produce better diversity among the learner base. RoDRF applies the variable rotation concept to trees based on the DRF algorithm. Random rotations or transformations on different feature subspaces produce different projections, leading to better generalization or prediction performance. This research aims to compare the performance of RoDRF with RF, RoF, and DRF models on imbalanced data in cases of food insecurity. Class imbalance will be handled with two methods, namely EasyEnsemble and SMOTE-NC. The research results show that the DRF's model with EasyEnsemble techniques produces a model with the best performance among several algorithms tested. Even though the resulting accuracy is 0.62274 and the AUC value is 0.68501, the model can predict each class equally. All algorithms with EasyEnsemble treatment have average AUC values significantly different from each other based on statistical test results. This research also used SHAP to explain variables significantly contributing to the household's food insecurity status model.