Neighborhood search with heuristic-based feature selection for click-through rate prediction

Abstract

Click-through-rate (CTR) prediction is crucial in online advertising and recommender systems. Maximizing CTR has been a major focus, leading to the development of numerous models designed to capture implicit and explicit feature interactions. However, extracting both low-order and high-order interactions remains challenging, as irrelevant features can increase computational costs and reduce prediction accuracy. Feature performance may also vary across predictive models and fluctuate due to traffic changes, making efficient feature selection essential in live environments where both training and inference times are critical. Traditional filter-based feature selection methods often fail to dynamically identify the most impactful features. This paper introduces a greedy heuristic, called Neighborhood Search with Heuristic-based Feature Selection (NeSHFS), to enhance CTR prediction by iteratively refining the feature set. NeSHFS employs a grid-search-like strategy to identify and retain the most relevant features, effectively reducing dimensionality and computational costs. Comprehensive experiments on several public datasets validate this approach, demonstrating improved prediction performance and reduced training times.