<p>AdaBoost improves model performance by combining weighted base classifiers. However, using approximate sample weights training base classifiers reduces ensemble diversity, and many pairwise diversity measures ignore differences between samples. Redundant base classifiers in ensemble models are inevitable, and particle swarm optimization (PSO) can remove them but is susceptible to local optimality. To overcome these limitations, an improved AdaBoost based on ensemble diversity and multi-population collaborative PSO (AdaBoost-D-IMCPSO) is proposed. Firstly, to consider differences between samples when measuring ensemble diversity, a novel diversity measure is proposed based on redefined sample importance. Secondly, to enhance the diversity of the AdaBoost, a new training strategy is proposed. The newly trained base classifier retrains if the diversity between the classifier and the existing model is poor. Finally, to improve the ability of PSO to escape from local optimality, an improved multi-population collaborative PSO (IMCPSO) with three strategies is proposed. These strategies help particles escape from local optimality through information shared by populations and gradually increase population diversity. Experiments on 18 public datasets confirm AdaBoost-D-IMCPSO outperforms state-of-the-art algorithms, improving accuracy by up to 10.67% compared to the highest-performing counterpart during the five-fold cross-validation. Additionally, Wilcoxon’s signed rank test shows a significant difference between the proposed method and seven counterparts. By enhancing ensemble diversity and reducing model complexity, the proposed method achieves a highly accurate and diverse model with a compact size.</p>