Optimizing Deep Neural Networks for Water Quality Prediction Using Spider Monkey Optimization and Permutation Importance-Based Feature Selection: A Comparative Heuristic Approach

Abstract

In this study, a novel integration of Spider Monkey Optimization (SMO) with deep neural networks (DNNs) has been proposed for hyperparameter tuning in the context of water potability classification. To the best of the authors’ knowledge, this represents one of the first systematic applications of SMO for this task. The proposed framework incorporates a comprehensive data preprocessing pipeline, which includes missing value imputation using a K-Nearest Neighbors (KNN) imputer, feature normalization through min–max scaling, and class balancing via the Synthetic Minority Oversampling Technique (SMOTE). The SMO optimized DNN has been benchmarked against three established metaheuristic algorithms, Particle Swarm Optimization (PSO), GreyWolf Optimizer (GWO), and the Firefly Algorithm (FA), under a uniform Receiver Operating Characteristic Area Under the Curve (ROC-AUC) based objective function and hyperparameter search space. The SMO-DNN yielded competitive performance, attaining an average accuracy of 0.6947 and a ROC-AUC score of 0.6955 under 10-fold cross-validation, while reducing total computation time to 1347.41 minutes, which is 35% faster than GWO. Permutation Feature Importance (PFI) analysis identified sulfate, pH, and hardness as the most influential features. Further training using this reduced feature set led to improvements in both F1-score and ROC-AUC, highlighting the robustness and efficiency of the approach. Overall, the proposed method provides a scalable and computationally efficient strategy for deep learning based water quality assessment and demonstrates the potential of SMO as a viable alternative for deploying interpretable and accurate potability classification models, particularly in resource-constrained monitoring environments.