摘要： 多示例学习是继监督学习、非监督学习、强化学习后的又一机器学习框架。将多示例学习和非监督学习结合起来，在传统非监督聚类算法K-means的基础上提出MI_K-means算法，该算法利用混合Hausdorff距离作为相似测度来实现数据聚类。实验表明，该方法能够有效揭示多示例数据集的内在结构，与K-means算法相比具有更好的聚类效果。

关键词: 多示例学习, K-means聚类, 包间距, 聚类有效性评价

Abstract: Multi-instance learning is a new machine learning framework following supervised learning, unsupervised learning and reinforcement learning. Multi-instance learning and unsupervised learning are combined. This paper proposes a new multi-instance clustering algorithm MI_K-means based on traditional unsupervised learning algorithm K-means. The algorithm MI_K-means adopts mixed Hausdorff distance as similar measure to carry out clustering. Experimental shows that MI_K-means can effectively reveal inherent structure of a multi-instance data set, and it can get better clustering effect than K-means algorithm.

Key words: multi-instance learning, K-means clustering, distance between bags, validity measure on clustering

中图分类号: 

• TP311