<p>Multi-source heterogeneous defect prediction typically faces challenges including feature space mismatch, class imbalance, and difficulty in optimizing classifier parameters. To address these issues, this paper proposes a multi-source heterogeneous defect prediction model based on SVM optimized by the filtered mutation projection ivy algorithm (FMPIVY-SVM). First, a transfer learning model is employed to align the feature spaces across different data sources. Second, to alleviate the class imbalance problem, a dynamic neighbor selection oversampling algorithm based on JS divergence (JSO) is proposed. Subsequently, by incorporating a Kalman filter update strategy and a mutation projection learning strategy, the filtered mutation projection ivy algorithm (FMPIVY) is introduced to dynamically optimize the classifier parameters and avoid local optima. Experimental results on performance evaluation demonstrate that FMPIVY achieves superior convergence speed and optimization accuracy compared to other benchmark algorithms. Finally, the optimized SVM is adopted as the classifier to construct the proposed multi-source heterogeneous defect prediction model. In comparative experiments with different optimization algorithms, the proposed model achieves mean F-measure, AUC, and G-mean values of 0.7331, 0.7972, and 0.7123, respectively, outperforming the other seven comparison models. This indicates that FMPIVY yields superior predictive performance in multi-source heterogeneous defect prediction. Experimental results on cross-project defect prediction show that FMPIVY-SVM significantly outperforms the other five comparative models in terms of both F-measure and AUC, demonstrating superior predictive capability.</p>