多示例学习的可行域定位及快速因果实例选择  

Feasible region localization and fast causal instance selection for multi-instance learning

作　　者：杨梅[1,2,3] 柯文静 王丹东[1] YANG Mei;KE Wenjing;WANG Dandong(School of Computer Science,Southwest Petroleum University,Chengdu 610500,Sichuan,China;Institute for Artificial Intelligence,Southwest Petroleum University,Chengdu 610500,Sichuan,China;Lab of Machin Learning,Southwest Petroleum University,Chengdu 610500,Sichuan,China)

机构地区：[1]西南石油大学计算机科学学院,四川成都610500 [2]西南石油大学人工智能研究院,四川成都610500 [3]西南石油大学机器学习研究中心,四川成都610500

出　　处：《山东大学学报（理学版）》2023年第9期105-113,126,共10页Journal of Shandong University(Natural Science)

基　　金：国家自然科学基金资助项目(62006200);四川省自然科学基金资助项目(2019YJ0314);浙江省海洋大数据挖掘与应用重点实验室开放课题(OBDMA202102);南充市校合作项目(SXHZ051)。

摘　　要：提出了一种多示例学习的可行域定位及快速因果实例选择(feasible region localization and fast causal instance selection for multi-instance learning,FFCM)算法,包含3个技术。可行域定位技术基于距离度量,从正包中选出具有代表性的实例作为候选实例;然后利用概率分析筛选负裁判包,以最大限度缩减选择因果实例的可行域范围。快速因果实例选择技术利用候选实例与负裁判包的因果关系构建融合包,设计因果性评判指标,使用先验知识从候选实例中选择出因果实例。包映射技术基于因果实例和差值映射函数,将包映射为有较高可区分性的单向量。本算法在27个常用数据集上进行了实验,并与6个前沿的MIL算法进行了对比,实验结果展示了FFCM的良好分类性能。This paper proposes a feasible region localization and fast causal instance selection(FFCM)algorithm for multi-instance learning,incorporating three techniques.To minimize the feasible region of data,the fast feasible region localization technique is used to select representative instances from the positive bags as candidate instances based on distance measurement,and reduces the negative referee bags through probability analysis.The fast causal instance-based selection technique uses the causal relationship between candidate instances and negative referee bags to construct fusion bags.Subsequently,prior knowledge is employed to select causal instances from candidate instances based on the designed causal instance criteria.The bag mapping technique maps bags into single vectors with high distinguishability using causal instances and a difference-based mapping function.The proposed algorithm is compared with 6 state-of-the-art MIL algorithms on 27 commonly used datasets.The experimental results show that the proposed FFCM exhibits comparable classification performance.

关 键 词：因果实例 可行域 映射 多示例学习 概率分析 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]