Data-Driven Feature Selection and Prediction of Municipal Waste Generation: Towards Sustainable Waste Management and Circular Economy Planning in the Slovak Republic

This study evaluates the performance of six feature selection methods (BORUTA, LASSO, RFE, XGBoost, FSM, and SEV) and five predictive modelling techniques (ANN, MARS, RST, SRT, and SVM) for the spatial estimation of municipal waste accumulation rates across 79 districts in the Slovak Republic. Using a 2022 cross-sectional dataset comprising 45 socio-economic and demographic variables, the study focuses on spatial prediction for unseen districts rather than temporal forecasting. Feature selection results indicate that BORUTA, RFE, and XGBoost consistently identify key predictors, notably the share of three-person households, the density of transport and warehousing companies, and average monthly wages. Model robustness was ensured through repeated random sub-sampling (30 iterations, 70/30 split) and validated using the Friedman test with Nemenyi post hoc comparisons (α = 0.05). The highest accuracy was achieved by MARS and ANN models coupled with SEV selection (MAE ≈ 28–30 kg/(person·year), MAPE ≈ 6%, R 2 > 0.88), and by SVM with XGBoost (MAE ≈ 30 kg/(person·year), R 2 ≈ 0.90). Reducing the predictor set from ten to five resulted in only minor performance degradation (MAPE increase < 1 pp), confirming the effectiveness of dimensionality reduction. The proposed approach enables accurate, computationally efficient waste generation estimation, thereby supporting regional planning and evidence-based policy development. In a broader context, the findings contribute to the implementation of the European Green Deal and circular economy objectives by providing tools for spatially targeted waste management strategies, directly aligned with United Nations Sustainable Development Goal 11 (Sustainable Cities and Communities) and Goal 12 (Responsible Consumption and Production).