部分对应的多视角聚类研究

Multi-view clustering integrates multiple information and achieves better results than single-view clustering. In the field of unsupervised learning, multi-view clustering becomes a hot topic in recent years. At present, multi-view clustering has the following problems: ① Most multi-view clustering algorithms only apply to completely mapped data, but the views tend to be partially mapped in a large number of applications. Only a small number of algorithms can be used for multi-view data with partially mapped instances but are not perfect. However, there have been no research on multi-view clustering on data with both partially mapped instances and clusters. ② Existing algorithms use only consistency or difference to interact among the views but do not consider the two aspects simultaneously. ③ Each view has different effects on information fusion, but there are few studies on the weighting scheme. More importantly, the existing research does not consider the weight of cluster. To solve the above problems, the project studies ① how to utilize consistency and difference simultaneously for completely mapped data; ② how to interact among the views when instances are partially mapped; ③ how to interact among the views when both instances and clusters are partially mapped; ④ how to weight the importance of each cluster in the process of integrating multi-view information. The project distinguishes various types of information from different views and uses the mapped information to connect multiple views, which can improve the effectiveness of multi-view clustering, expand the scope of multi-view clustering applications, and contribute a lot for perfecting the theory of multi-view learning as well as solving practical problems in related fields.

多视角聚类融合多个视角的信息，取得比单视角聚类更好的效果，是近年来无监督学习的研究热点。目前多视角聚类研究还存在以下问题：①大多数算法仅能用于完全对应的视角，但多数应用中视角通常只能部分对应。目前仅有少量算法可用于部分对应的数据实例，尚不完善。而数据实例和簇均部分对应的情况还没有人研究。②现有算法仅利用视角间一致性或差异性，没有同时考虑两个方面。③ 每个视角对信息融合的作用不同，但关于视角加权的研究很少。更重要的是，现有研究没有考虑簇权重。针对上述问题，本项目研究①如何同时利用一致性和差异性聚类完全对应数据；②如何利用部分对应的数据实例传递视角信息；③如何利用部分对应的数据实例和簇类别传递视角信息；④如何在融合多视角时体现簇权重。本项目区别各视角的各类信息并使用视角间对应信息连接多视角，可提升多视角聚类效果、扩大多视角聚类应用范围，为完善多视角学习理论、解决相关领域的实际问题做出重要贡献。

大多数多视角聚类算法假设数据实例和簇是完全对应的，然而，实际应用中的多视角数据可能是部分对应的，在某些视角中，数据实例和簇可能会丢失。此外，每个视角对数据聚类贡献的有效信息量不同，在融合多个视角的信息时，有必要区别每个视角，最大化利用每个视角特征空间的信息。因此，本项目在完全多视角聚类、部分多视角聚类、加权多视角聚类和多视角应用等方面展开研究，主要创新成果包括：（1）在完全多视角聚类方面，同时挖掘视角间的一致性和差异性，通过自监督学习策略充分利用每个视角的信息，利用交叉重构的方式融合多个视角的信息，设计端到端的策略获得聚类结果，提升了多视角聚类的效果；（2）在部分多视角聚类方面，提出基于谱聚类的数据实例部分对应的多视角聚类方法和基于非负矩阵分解的数据实例和簇部分对应的多视角聚类方法，扩展了多视角聚类的应用范围；（3）在加权多视角方面，提出赋簇权重的多视角聚类算法；（4）在多视角应用方面，将多视角思想应用于跨模态检索、小样本学习、约束聚类领域。本项目在 Neural Networks、Knowledge-Based Systems等国际期刊和 SIGIR、CIKM、ICME 等国际会议上发表论文 13 篇，申请国家发明专利1项。