【摘要】 糖尿病特征具有多模态、高维度、冗余复杂等特点，影响糖尿病的预测精度。针对此问题，本文提出基于特征交互与改进多目标粒子群优化算法的糖尿病特征选择方法。首先，对多模态数据进行融合，利用文本与数值数据共同构建糖尿病特征集；然后，通过mRMR与改进多目标粒子群优化算法删除冗余特征，筛选出与糖尿病相关性高的重要特征；最后，基于相关性分析构建组合指标特征集。采用SVM、随机森林、逻辑回归和决策树等4种预测模型对其进行分类评估，结果显示组合指标特征集的预测准确率为90%。对糖尿病预测有较高的准确率。更多还原

【Abstract】 The paper proposes a diabetes feature selection method based on feature interaction and improved multi-objective particle swarm optimization algorithm to address the characteristics of multimodality, high dimensionality, redundancy, and complexity in diabetes prediction. Firstly, multimodal data is fused to construct a diabetes feature set using both textual and numerical data. Then, redundant features are eliminated through mRMR and the improved multi-objective particle swarm optimization algorithm to select important features highly correlated with diabetes. Finally, a composite indicator feature set is constructed based on correlation analysis. Four prediction models including SVM, Random Forest, Logistic Regression and Decision Tree are employed for classification evaluation. The results show that the prediction accuracy of the composite indicator feature set reaches 90%, indicating a high accuracy in diabetes prediction.更多还原