基于不完整标签的增强低秩表示用于预测阿尔茨海默病进展

Enhanced Low-Rank Representation with Incomplete Labels for Predicting Alzheimer’s Disease Progression

摘要: 本文提出了一种名为基于不完整标签的增强低秩表示(ELRRIL)模型，用于解决神经影像数据中不完整标签样本和噪声问题，进而提高阿尔茨海默病(AD)进展预测的准确性。我们的方法利用矩阵分解技术，将不完整的认知评分矩阵分解为两个组成部分：一方面，通过增强的流形正则化恢复的无缺失值的认知评分矩阵，该正则化能够捕捉局部标签相关性；另一方面，基于噪声的稀疏假设，通过ℓ₁范数控制的错误分量。最后，我们使用低秩回归模型，将恢复的矩阵作为目标，提高对噪声和异常值的鲁棒性，并引入了ℓ_2,1范数作为稀疏正则化项来识别重要的神经影像特征。实验结果表明，ELRRIL模型在特征选择和预测性能方面均优于现有的先进方法。

Abstract: In this paper, a new model called Enhanced Low Rank Representation with Incomplete Labels (ELRRIL) is proposed to solve the problem of incomplete label samples and noise in neuroimage data, thereby improving the accuracy of predicting the progression of Alzheimer’s disease (AD). Our method uses matrix decomposition techniques to decompose the incomplete cognitive score matrix into two components: one is the missing value-free cognitive score matrix recovered by enhanced manifold regularization, which can capture local label correlations; The other is the error component controlled by the ℓ₁ normbased on the sparse assumption of noise. In addition, we develop a low-rank regression model that targets the recovered matrix to improve robustness to noise and outliers, and introduce the ℓ_2,1 norm as a sparse regularization term to identify important neuroimage features. Experimental results show that the ELRRIL model is superior to the existing advanced methods in feature selection and prediction performance.