IMPLEMENTATION OF CUCKOO SEARCH-BASED ENSEMBLE VARIABLE IMPORTANCE IN THE CLASSIFICATION OF NON-CASH FOOD ASSISTANCE (BPNT) RECIPIENTS IN WEST JAVA

Abstract

The BPNT program is a government initiative to efficiently distribute social assistance to poor households. However, the challenge of achieving accurate recipient identification remains a major obstacle. This research aims to build a classification model for BPNT recipients in West Java using machine learning methods (Random Forest, XGBoost, CatBoost, and LightGBM) and a Cuckoo Search-Based Ensemble Variable Importance (EVI) approach to identify which predictors most strongly affect classification. Class imbalance in the response data was addressed through weighting during model training, and performance was evaluated using balanced accuracy through 10-fold cross-validation. Although all models performed well, the variable importance results varied across models. Using the Random-Key Cuckoo Search algorithm, an EVI ranking was generated that integrated VI rankings from each model, achieving a minimum Spearman correlation of 0.6538. The results show that roof quality, living status, calorie consumption, and per capita expenditure are the main indicators for classifying BPNT recipients. This approach shows great potential to improve modeling interpretability and to provide stronger data-driven support for social policy-making.