Jing Ling, Li Yi, Zhang Hongjie
College of Science, China Agricultural University, Beijing, China.
College of Information and Electrical Engineering, China Agricultural University, Beijing, China.
PeerJ Comput Sci. 2024 Dec 20;10:e2619. doi: 10.7717/peerj-cs.2619. eCollection 2024.
Feature extraction research has witnessed significant advancements in recent decades, particularly with single-view graph embedding (GE) methods that demonstrate clear advantages by incorporating structural information. However, multi-view data includes descriptions from various perspectives or sensors, offering richer and more comprehensive information compared to single-view data. Research interest in multi-view feature extraction is steadily increasing. Hence, there is a pressing need for a comprehensive framework that extends single-view methods, especially effective GE methods, into multi-view approaches. This article proposes three innovative multi-view feature extraction frameworks based on regression embedding. These frameworks extend single-view GE methods to the multi-view scenario. Our approach meticulously considers the consistency and complementarity of multi-view data, placing strong emphasis on robustness to noisy datasets. Additionally, the use of non-linear shared embedding helps prevent the loss of essential information that may occur with linear projection techniques. Through numerical experiments, we validate the effectiveness and robustness of our proposed frameworks on both real and noisy datasets.
近几十年来，特征提取研究取得了显著进展，特别是单视图图嵌入（GE）方法，通过纳入结构信息展现出明显优势。然而，多视图数据包含来自不同视角或传感器的描述，与单视图数据相比，提供了更丰富、更全面的信息。对多视图特征提取的研究兴趣正在稳步增加。因此，迫切需要一个全面的框架，将单视图方法，特别是有效的GE方法扩展到多视图方法中。本文提出了三种基于回归嵌入的创新多视图特征提取框架。这些框架将单视图GE方法扩展到多视图场景。我们的方法精心考虑了多视图数据的一致性和互补性，特别强调对噪声数据集的鲁棒性。此外，使用非线性共享嵌入有助于防止线性投影技术可能导致的重要信息丢失。通过数值实验，我们在真实数据集和噪声数据集上验证了所提出框架的有效性和鲁棒性。
