Intelligent automated feature engineering and modelOptimization using reinforcement learning for large-scalepredictive analytics applications

Abstract

Feature engineering, feature selection, model selection and hyperparameter tuning are typically large and manual tasks that can be costly in terms of computational resources, time to develop and suboptimal predictive performance in large-scale predictive analytics applications. Typically, conventional automatic machine learning technologies can minimize the need for manual participation, but often require a pre-designed search strategy that may not be able to adjust well to complex and evolving optimization spaces. In this work, these drawbacks are overcome by proposing an intelligent reinforcement-learning-based framework for the simultaneous task of feature engineering and predictive model optimization. The proposed approach treats the feature transformation, feature selection, model selection, and hyperparameter configuration as sequential decision-making problems, allowing a reinforcement learning agent to explore different combinations of them and learn optimal strategies based on the reward feedback when they are performed. The proposed framework combines automated feature generation and selection with adaptive model and hyperparameter optimization to find optimal predictive configurations that balance predictive performance and computational efficiency. The framework is tested with big datasets of predictive analytics and is contrasted to traditional feature engineering, standard machine learning optimization, and automated optimization methods. The measure of the performance depends on the prediction task, and it is evaluated by predictive accuracy, precision, recall, F1-score, RMSE, computational time, convergence rate, and resource utilization. Experimental results show that reinforcement learning is effective in reducing unnecessary combinations of features, as well as in tuning better models and selecting them efficiently, while maintaining competitive prediction performance with lower optimization costs. The main contribution is an integrated intelligent optimization framework, combining automated feature engineering and model optimization with reinforcement learning, enabling the efficient predictive analytics on large and complex datasets.