多视图作为多视角数据的一种表现形式，被广泛应用在有监督、半监督和无监督学习中。当前大部分基于多视图的学习方法主要聚焦于视角数据之间的互补性，忽略了视角数据之间的差异性、模型的鲁棒性和可解释性。为此，本项目拟在协同学习方法支撑下，以子空间学习、多核学习和非负矩阵分解等为相关技术，循序渐进地按照“多视图恢复”、“鉴别多视图差异提升”、“多视图约束残差保持”和“自适应多视图嵌入”等思路开展鲁棒多视图协同学习理论与方法研究；应用协同学习方法，模型有效融合和提升多视图之间的互补性和差异性来发现数据的潜在结构；通过引入稀疏约束以及自适应视角图权重学习，使模型能自动识别重要视图数据及自动发现多视角数据的潜在模式，增强模型的可解释性；本项目的研究对完善协同学习、多视角学习和图学习理论体系具有重要的学术意义，其研究成果在协同计算、数据挖掘、计算机视觉和模式识别等都有重要的理论价值和广泛的应用前景。

Multi-view graph which is used as a representation form of multi-view data, is commonly used in supervised, semi-supervised, and unsupervised learning scenarios. Currently, most of multi-view graph based learning methods mainly focus on the complementarity of multi-view data. However, the diversity of multi-view data and robustness and interpretability of model are commonly ignored. To this end, under the idea of collaborative learning, this research project gradually proposes the ideas of “multi-view graph recovery”, “diversity promoting of discriminant multi-view graph”, “multi-view graph constraint multiple residual preserving” and “adaptive multi-view graph embedding” to develop new theory and method on robust collaborative multi-view graph learning by introducing subspace learning, non-negative matrix factorization and multiple kernel learning. By using the idea of collaborative learning, models effectively merge and promote the complementarity and diversity of multi-view graphs to discover the underlying structure of data. By introducing the sparse constraint and adaptive multi-view graph weights learning, models can automatically identify the most multi-view graph data and discover the underlying model of multi-view data, enhancing the interpretability of model. This project has important academic significance to enrich the collaborative learning, multi-view learning and graph learning theoretical system. The related theories and algorithms can be widely used in collaborative computing, data mining, computer vision and pattern recognition and so on.