In the era of increasing Industrial Internet of Things (IIoT) devices, effective and efficient anomaly detection in network traffic is crucial for ensuring the security and reliability of industrial systems. This paper introduces a systematic methodology for optimizing machine learning models by focusing on the critical trade-off between detection accuracy and computational efficiency for resource-constrained IIoT environments. The methodology was evaluated using decision tree-based algorithms (RandomForest, ExtraTrees, AdaBoost, XGBoost, CatBoost) on a realistic dataset with simulated network attacks. The analysis involved a comprehensive evaluation of data preparation strategies, including class balancing, data aggregation, sampling, feature selection, and hyperparameter tuning, with a specific focus on the XGBoost model. The results demonstrate that this holistic optimization enables high detection accuracy (over 92% for binary classification and 87% for multi-class classification) while simultaneously maintaining high computational efficiency (short training time, small model size). This approach provides a practical pathway for developing resilient and resource-aware cybersecurity systems for industry, smart city, and IIoT environments.