基于实例的多视角多标签学习算法

Multi-View Multi-Label Learning by Exploiting Instance Correlations

摘要: 在多视角多标签学习中，一个实例可以由多个数据视角表示，并与多个类标签相关联。目前现有的多视角多标签学习算法研究的是变量与类标签之间的关系。在这篇文章中，我们提出了一种基于实例的多视角多标签学习算法。与现有的多视角多标签算法不同，我们利用训练实例和测试实例的相关性进行建模。在多视角多标签学习中，数据来自多个视角。在每个视角中，通过构建偏最小二乘回归模型来探索训练实例和测试实例之间的相关性，而不是像传统的多视角多标签学习算法一样探索变量和类标签之间的映射函数。因此，我们可以从多视角数据中学习到多个偏最小二乘回归模型。此外，为了保证不同视角的一致性，在多个回归模型中加入视角一致性约束。对比实验表明，与现有的多视角多标签算法相比，所提算法获得了更好的分类结果。

Abstract: In multi-view multi-label learning, an instance can be represented by multiple data views, and associated with multiple class labels. The existing multi-view multi-label learning works explore the relationship between the variables and the class labels. In this paper, we present an instance-based multi-view multi-label learning algorithm. Different from the existing multi-view multi-label algo-rithms, we model the correlations of training instances and testing instances. In multi-view multi-label learning, the data is from multiple views. In each view, we build a partial least square re-gression model to explore the correlation between the training instances and the test instances, rather than the mapping function between the variables and the class labels. As a result, we can learn multiple partial least square regression models from multi-view data. Moreover, in order to ensure the mutual agreement on distinct views, a view consensus constraint is proposed based on the multiple regression models. The comparison experiments show that our algorithm obtains better classification results in comparing with the state-of-the-art multi-view multi-label algorithms.