Correlation discovery for multi-view and multi-label learning

• Online Access: http://hdl.handle.net/2047/D20409219

Correlation indicates the interactions or connections across different instances. It exists in a wide range of real-world scenarios such as scene understanding, social network, time-series data, and human-object interactions. Correlation provides the unique and informative knowledge to reveal the latent connections across instances, and it plays an essential and important role in the machine learning field. However, recovering and utilizing correlation is challenging. First, it is hard to explicitly and clearly define the correlations, which leads to relatively small and high-level noise datasets. Second, the correlation is task-specific, which cannot be generalized to more diverse tasks. This challenge increases the cost of correlation learning and its down-stream applications. Third, even if the correlation is given, how to efficiently utilize the learned/given correlations and enhance the final performance is still difficult. This point has not been well-explored. In this dissertation research, we investigate the techniques to effectively discover various kinds of correlations in machine learning tasks including multi-view learning, multi-label learning, image/scene understanding, time-series data analysis, and human action recognition. Specifically, we propose algorithms from the following perspectives: 1) designing correlation exploration frameworks to automatically explore the label correlations in multi-label scenarios, 2) proposing a multi-view fusion strategy which effectively dig the latent correlations across different views to achieve high-accuracy human action recognition, and 3) exploring the inductive and unsupervised graph representation learning task, which aims to preserve the correlation knowledge in graph structured objects. To demonstrate the effectiveness of the proposed algorithms, various experiments on commonly used datasets have been implemented and the results show the superiority of our algorithms over the other state-of-the-art methods.--Author's abstract