一种基于示例非独立同分布的多示例多标签分类算法

多示例多标签学习是一种新型的机器学习框架。在多示例多标签学习中,样本以包的形式存在,一个包由多个示例组成,并被标记多个标签。以往的多示例多标签学习研究中,通常认为包中的示例是独立同分布的,但这个假设在实际应用中是很难保证的。为了利用包中示例的相关性特征,提出了一种基于示例非独立同分布的多示例多标签分类算法。该算法首先通过建立相关性矩阵表示出包内示例的相关关系,每个多示例包由一个相关性矩阵表示；然后建立基于不同尺度的相关性矩阵的核函数；最后考虑到不同标签的预测对应不同的核函数,引入多核学习构造并训练针对不同标签预测的多核SVM分类器。图像和文本数据集上的实验结果表明,该算法大大提高了多标签分类的准确性。

Multi-instance Multi-label Learning Algorithm by Treating Instances as Non-independent Identically Distributed Samples

Multi-instance multi-label learning (MIML) is a new machine learning framework.In this framework,an object is represented as a bag which is decomposed of multiple instances and labeled with multiple labels.Previous studies on MIML typically treated instances in the bags are independently identically distributed.However,it is difficult to be guaranteed in real tasks.Considering correlation features of instance in a bag,a multi-instance multi-label learning algorithm by treating instances as non-independent identically distributed samples was proposed.Firstly,instance correlations are considered by establishing an affinity matrix.By this means each bag is represented with an affinity matrix.Then,kernel functions based on the affinity matrix in different scales are established.Finally,considering predictions of different kinds of labels are corresponding to different kernels,multiple kernel learning is introduced to construct and train the MKSVMs.Experimental results on image data set and text data set show that the proposed algorithm greatly improves the accuracy of the image multi-label classification compared with other methods.