一种基于多示例学习的动态样本集半监督聚类算法

A Semi-supervised Dynamic Sample Set Clustering Algorithm Based On Multi-instance Learning

针对时域空间动态样本模式分类和标记信息的有效利用问题,提出了一种基于多示例学习的动态样本半监督聚类算法。根据时间信号的结构关系和模态特征,建立动态样本的多示例信息表示模型;通过定义一种可度量时变函数样本包间近似度的广义Hausdorff距离和基于近邻传播的聚类原则,构建多示例动态样本包的半监督聚类算法。算法利用样本包的类别先验知识构建样本集初始划分种子簇并探索样本的分布特征,采用基于广义Hausdorff距离的近邻传播策略调整样本包聚类,突出动态指标局部模态变化特征在样本分类中的差异性。以油田地质研究中测井曲线油层水淹状况判别为例,验证了模型和算法的有效性。

Aiming at the effective use of time-domain space dynamic sample's pattern classification and mark information in time-domain space,a multi-instance learning-based semi-supervised dynamic sample clustering algorithm was proposed. Basing on both structural relationship and modal characteristics of time signals,the multi-instance information model for dynamic samples was established; through defining a generalized Hausdorff distance which measures the similarity among time-varying function samples and considering affinity propagation-based clustering principle,a semi-supervised clustering algorithm for multi-instance dynamic samples was founded. This algorithm adopts category priori knowledge to build sample set's initially-partitioned seed cluster and to explore samples' distribution characteristics; and it adjusts sample clustering dynamically by adopting the generalized Hausdorff distance-based affinity propagation strategy so as to highlight dynamic index modal feature's individual difference in sample classification. Taking the recognition of oil layer's water flooded condition in well logging as an example,both model and algorithm's effectiveness was proved.