【摘要】 多示例学习中的数据是由包含多个示例的包所组成的，样本之间真实的相似性被正包中大量的假正例所掩盖。为了拟合多示例数据真实的分布情况，提出了一种结合Fisher编码的多示例聚类算法MIFK-means。首先通过Fisher编码在保留数据语义的同时对多示例数据进行归一化降维，然后通过示例层次的K-means聚类算法揭示多示例数据的分布情况。在基准数据集上的实验表明，MIFK-means算法的聚类效果明显好于包层次的多示例K-means聚类算法，分类精度也可以和现有的经典多示例算法相媲美。更多还原

【Abstract】 The samples of multi-instance learning is composed of bags containing multiple instances. The true similarity between bags is concealed by a large number of false positive instances in positive bags. In order to fit the true distribution of multi-instance data set, a multi-instance1 clustering algorithm combined with Fisher coding, i.e., MIFK-means is proposed. This method first uses Fisher coding to normalize and reduce the dimension of multi-instance data while preserving the data semantics, then the instance level K-means clustering algorithm is utilized to reveal the distribution of multi-instance data. Experiments on Musk data sets show that the clustering performance of MIFK-means algorithm is significantly better than that of bag level multi-instance K-means algorithm, and the classification accuracy of MIFK-means algorithm is comparable to the stat-of-the-art multi-instance algorithms.更多还原