Multi-instance clustering with applications to multi-instance prediction
Multi-instance clustering with applications to multi-instance prediction
复制标题
DOI：
10.1007/s10489-007-0111-x
复制
发表时间：
2009-08
影响因子：
5.3
通讯作者：
Min-Ling Zhang;Zhi-Hua Zhou
中科院分区：
文献类型：
--
作者：
Min-Ling Zhang;Zhi-Hua Zhou
在多实例学习设置中，每个对象由多个实例组成的abg来表示，而不是传统学习设置中的单个实例。这个领域以前的工作只关注多实例预测问题，其中每个包与一个二值（分类）或实值（回归）标签相关联。然而，无监督的多实例学习还没有被研究过。本文针对无监督多实例学习问题，提出了一种多实例聚类算法bamicis。简而言之，通过将袋视为原子数据项，并使用某种形式的距离度量来度量袋之间的距离，bamici采用了流行-介质算法将未标记的训练袋划分为不连接的袋组。在聚类结果的基础上，提出了一种新的多实例预测算法bartmipis。首先，将每个袋子用ak维特征向量重新表示，其中第i个特征的值设为袋子与第i组的中间点之间的距离。然后，将袋子转换为特征向量，以便使用普通监督学习器从转换后的特征向量中学习，每个特征向量与原始袋子的标签相关联。大量的实验表明，bamic1可以有效地发现数据集的底层结构，bartmip在各种多实例预测问题上都有很好的效果。
In the setting of multi-instance learning, each object is represented by abagcomposed of multiple instances instead of by a single instance in a traditional learning setting. Previous works in this area only concern multi-instancepredictionproblems where each bag is associated with a binary (classification) or real-valued (regression) label. However,unsupervisedmulti-instance learning where bags are without labels has not been studied. In this paper, the problem of unsupervised multi-instance learning is addressed where a multi-instance clustering algorithm namedBamicis proposed. Briefly, by regarding bags as atomic data items and using some form of distance metric to measure distances between bags,Bamicadapts the populark-Medoidsalgorithm to partition the unlabeled training bags intokdisjointgroups of bags. Furthermore, based on the clustering results, a novel multi-instance prediction algorithm namedBartmipis developed. Firstly, each bag is re-represented by ak-dimensional feature vector, where the value of thei-th feature is set to be the distance between the bag and the medoid of thei-th group. After that, bags are transformed into feature vectors so that common supervised learners are used to learn from the transformed feature vectors each associated with the original bag’s label. Extensive experiments show thatBamiccould effectively discover the underlying structure of the data set andBartmipworks quite well on various kinds of multi-instance prediction problems.