面向多标签数据的自步流形学习方法研究

Feature representation for multi-label data is a fundamental problem in pattern recognition and machine learning. Manifold learning methods achieve feature representation by catching and preserving the intrinsic geometrical structures of data. Self-paced learning methods gradually involve from easy to more complex training samples into the learning process. This strategy is helpful in alleviating the local optimum problem in non-convex optimization. Considering the characteristics and application backgrounds of multi-label data, this project utilizes manifold and self-paced learning theories to achieve feature representation. In order to learn the feature representation for image classification, a convolutional neural network is investigated, which could express the deep features, manifold structures and label structures. This network can also improve the performance of image classification. In order to achieve the fast retrieval of massive-scale multi-label data, a binary coding model is investigated, which could preserve the similarity between data points. This model can also reduce the storage cost and improve the search accuracy. A self-expressive auto-encoder network which could express the manifold and sparse structures, as well as a theoretical approach for fusion the similarity between the data points from multiple modals, are both investigated to realize the unified representation of multi-modal multi-label data. In addition, some training samples selection methods using self-paced learning theory are investigated, to reduce the influence of noises and outliers, as well as to improve the robustness and generalization ability. This project will not only enrich the theoretical systems and implementation approaches of manifold learning, self-paced learning as well as multi-label learning, but also paves good ways for image classification, data retrieval and information fusion.

多标签数据的特征表示是模式识别与机器学习中的基础问题。流形学习通过捕获与保持数据内在几何结构实现特征表示。自步学习从易到难学习数据样本，避免非凸模型陷入局部最优解。项目基于流形学习与自步学习理论，针对多标签数据的特点与应用背景开展特征表示研究。研究结合深度特征、流形结构与标签结构的卷积网络模型，解决图像多标签数据分类中的特征表示问题，并提高分类效果；研究保持数据语义相似度的二值编码模型，解决大规模多标签数据的快速检索问题，并降低存储代价、提高检索精度；研究能同时体现流形结构与稀疏结构的数据自编码网络模型，并研究融合不同模态数据相似度的理论方法，解决多模态多标签数据的统一表示问题；利用自步学习理论研究训练样本的选择方法，降低噪声与孤立点的影响，提升模型鲁棒性与泛化能力。研究成果不仅可丰富流形学习、自步学习与多标签学习的理论体系和实现途径，而且为图像分类、数据检索、信息融合等研究奠定基础。

按项目申报书要求，本项目围绕多标签数据的表示学习方法与学习机制开展研究，主要研究图像多标签数据分类的特征表示方法、多标签数据紧凑编码方法、多视角多标签数据的融合表示方法。拟在国际高水平期刊与会议发表论文5-6篇，其中SCI论文4篇以上，并且CCF B类（或中科院2区）以上刊物论文2篇以上，申请发明专利1-2项。培养硕士研究生2-3名，协助培养青年教师1-2名。三年来，项目组成员围绕既定目标进行深入研究并取得较大科研进展。具体而言，在图像表示方面，提出基于低秩标签分解的多标签分类算法，有效提高图像表示质量；在紧凑编码方面，提出面向多标签数据分类的结构特征表示算法，显著提高数据分类精度；在多视角数据融合表示方面，提出深度子空间相似融合算法与深层子空间互学习算法，有效提升多视角数据融合效果。基于上述研究内容，项目共发表论文7篇，其中包含Bioinformatics、IEEE/ACM Transactions on Computational Biology and Bioinformatics、Neurocomputing等著名SCI论文5篇，其中CCF B类期刊2篇，申请发明专利2项，指导硕士研究生4人，协助培养青年教师1人，已完成项目计划书制定的所有目标。项目负责人目前正积极与企业开展合作，有望将项目研究成果进行试点与推广。