基于特权信息学习的多视图支持向量机

Multi-View Support Vector Machine Based on Privileged Information Learning

摘要: 多视图支持向量机是处理多视图特征数据的典型学习方法，在小样本、高可解释性需求场景下仍具有不可替代的优势。现有多视图支持向量机方法直接在原始高维数据上构建分类器，未考虑高维数据中的冗余与噪声，且多数将“降维”与“分类”视为独立过程，导致分类性能受限。针对上述问题，本文提出一种基于特权信息学习的多视图支持向量机(MVDRP)，首次将高维数据降维与多视图SVM分类器训练融合为统一凸优化模型并同步优化，同时引入特权信息学习实现视图间互补信息的双向传递，确保模型同时满足多视图学习的一致性与互补性原则。在图像、文本、多媒体三类真实多视图数据集上的实验表明，本文模型不仅在分类精度与F1分数上优于传统“降维 + 分类”串联方法及现有多视图SVM方法，在小样本、低计算资源场景下相比主流深度学习多视图方法也具有显著优势。

Abstract: Multi-view Support Vector Machine is a typical learning method for processing multi-view feature data, which still has irreplaceable advantages in scenarios requiring small samples and high interpretability. Existing MvSVM methods construct classifiers directly on original high-dimensional data without considering the redundancy and noise, and most treat “dimensionality reduction” and “classification” as independent processes, leading to limited classification performance. To address these issues, this paper proposes a Privileged Information Learning-based Multi-view Support Vector Machine. For the first time, this model integrates high-dimensional data dimensionality reduction and MvSVM classifier training into a unified convex optimization model for simultaneous optimization. Meanwhile, privileged information learning is introduced to realize bidirectional transmission of complementary information between views, ensuring that the model conforms to both the consistency and complementarity principles of multi-view learning. Experiments on three types of real-world multi-view datasets (image, text and multimedia) show that the proposed model outperforms traditional “dimensionality reduction + classification” tandem methods and existing MvSVM methods in both classification accuracy and F1-score, and also has significant advantages over mainstream deep learning multi-view methods in small sample and low computational resource scenarios.