In recent years, Graph Neural Networks (GNNs) have garnered significant attention, with a notable focus on Graph Structure Learning (GSL), a branch dedicated to optimizing graph structures to enhance network training performance. Current GSL methods primarily involve constructing optimized graph representations by analyzing one or more initial graph sources to improve performance in subsequent application tasks. Despite these advancements, achieving high-quality graphs that accurately and robustly reflect node relationships remains challenging. This paper introduces a novel approach, termed BAB-GSL, designed to approximate an ideal graph structure through a systematic process. Specifically, two basic views are extracted from the original graph and utilized as inputs for the model, where the preliminary optimized view is generated through the view fusion module. The Attention mechanism is then applied to the optimized view to improve nodes’ connectivity and expressiveness. Subsequently, the trained view is re-structured using a Bayesian optimizer to produce the final graph structure. Extensive experiments were conducted across multiple datasets, both in undisturbed and attacked scenarios, to thoroughly evaluate the proposed method, demonstrating the effectiveness and robustness of the BAB-GSL approach.