基于不确定比例标签信息的多视角学习

比例标签数据作为弱标签数据中的一种，近年来受到了广泛关注。针对现实应用中所获取的比例标签数据存在噪声污染而标注困难的局限性，提出将多视角学习应用到不确定比例标签学习中以提高分类器的准确率和稳定性。首先，所提出的方法利用核方法映射多视角信息到特征空间以及引入超实例标签表示比例标签信息，在将多视角信息与比例标签信息结合的同时将不确定比例标签数据的局限性考虑在内，以此构建基于不确定比例标签信息的多视角学习模型；而后利用拉格朗日方法将目标模型转换为对偶形式求解，并引入一个迭代框架解决目标模型以获取多视角分类器。最后，将所提出的方法应用于文本分类的数据集，实验结果表明提出的方法比已有的单一视角比例标签学习方法更具优越性，具体表现在所提方法更高的分类准确率和更不易受数据噪声影响等方面。

Uncertain label proportions information-based multi-view learning

Label proportions data, one of the weakly labeled data, has drawn extensive attention in recent years. It is difficult to label the label proportions data contaminated by noise in practical applications. This paper applied multi-view learning to uncertain learning with label proportions to improve classifier both performance and stability. First of all, the proposed method introduced kernel function and super instance label for the representation of multi-view knowledge and label proportions information respectively, and considered the limitation of uncertain label proportions data to construct the model of uncertain label proportions information-based multi-view learning. Then, it transformed the model to its dual forms by the Lagrange method for solving the problem and introduced an iterative framework to obtain the classifiers of multiple views. Afterward, this paper applied the proposed method to datasets of text categorization. Extensive experiments show that the proposed method outperforms the existing single-view learning with label proportions methods in accuracy and noise sensitivity.