Clustering-based classifier ensemble learning methods have been extensively studied for their proficiency in classifying noisy datasets in real-world scenarios. However, their application is often constrained by two significant limitations. Firstly, the common methodologies for generating data clusters typically rely on a fixed or derived value fro the number of clusters from the training data, which may not always be optimal. Secondly, the selection of classifiers from a pool, a task inherently combinatorial in nature, presents a challenge due to its NP-hard complexity. This makes the task of achieving an optimal balance between maximizing ensemble accuracy and minimizing the size of its components very challenging. To tackle these issues, we propose a novel clustering-based ensemble method. Our method starts by identifying the optimal number of clusters, unlike the conventional fixed or derived approaches. Next, we approach classifier selection as a discrete optimization problem and propose a modified binary Particle Swarm Optimization algorithm for selecting the most effective classifier subset. We thoroughly evaluate our algorithm using various benchmark datasets from the UCI machine learning repository, demonstrating enhanced performance and efficiency compared to both single classifiers and existing ensemble methods.