AutoField-POA: A Hybrid AutoML–Metaheuristic Framework for Automated Feature Selection in Recommender Systems

The problem of automated feature selection is one of the most important aspects of modern recommender systems, as overestimating the number of feature fields used, which can increase model complexity, inference latency, and performance degradation due to redundancy and noise. Although differentiable Automated Machine Learning (AutoML) based feature selection methods, including AutoField, are effective, they remain susceptible to initialization and local optima. Conversely, population-based metaheuristic algorithms are effective at global exploration, but are often inefficient in high-dimensional discrete spaces when randomly initialized. This paper presents AutoField-POA, a hybrid AutoML-metaheuristic framework that combines a differentiable controller with the Pelican Optimization Algorithm (POA) to solve automated feature selection problems. The controller guides population initialization toward promising areas of the search space, whereas POA focuses on feature subsets through a balanced exploration–exploitation search strategy. A composite fitness function is jointly optimized to achieve both predictive accuracy and subset compactness. Large-scale Click-Through Rate (CTR) experiments on the Criteo and Avazu datasets indicate that AutoField-POA consistently achieves better performance than AutoField and baseline methods in Area Under the ROC Curve (AUC), Logloss, efficiency, and stability, and exhibits high transferability across recommender architectures.