基于最大化交叉互信息的对称IB算法

发布时间：2018-08-31 21:00

【摘要】：对称IB(Symmetric Information Bottleneck)通过行、列压缩变量之间的相互协作来挖掘数据中的双向压缩模式.由于行、列压缩变量不能完全承载行、列基层变量中所蕴含的特征信息,从而导致对称IB所得的数据双向压缩模式与基层变量所蕴含的内在模式之间存在一定的偏离.针对该问题,通过最大化地保存压缩变量与基层变量交叉之间的互信息,将基层变量引入到数据的双向压缩中,使它们协助压缩变量共同来学习联合分布中的双向压缩模式,提出交叉对称IB:ICSIB(Inter-Correlated Symmetric Information Bottleneck).ICSIB算法采用交错的顺序"抽取-合并"迭代过程来优化压缩变量与基层变量交叉之间的互信息,可保证得到目标函数的一个局部优解.实验结果表明,在基层特征变量的协助下,ICSIB算法得到的数据双向压缩模式更接近于数据中真实的内在模式,并可有效地应用于数据的联合聚类中.
[Abstract]:Symmetric IB (Symmetric Information Bottleneck) mine bidirectional compression patterns in data by the collaboration of row and column compression variables. Because of the row, the column compression variable can not carry the row completely, and the characteristic information contained in the column base variable, which leads to the deviation between the bidirectional compression mode of symmetric IB data and the inherent pattern contained in the base variable. To solve this problem, by maximizing the mutual information between the compressed variables and the basic variables, the basic variables are introduced into the bidirectional compression of data, so that they can help the compressed variables to learn the bidirectional compression mode in the joint distribution. In this paper, a cross-symmetric IB:ICSIB (Inter-Correlated Symmetric Information Bottleneck). ICSIB) algorithm is proposed to optimize the mutual information between the compressed variables and the basic variables by using the interleaved sequential "decimation-merging" iteration process, which can guarantee a local optimal solution of the objective function. The experimental results show that the data bidirectional compression pattern obtained by ICSIB algorithm with the help of basic feature variables is closer to the real internal pattern and can be effectively applied to the joint clustering of data.
【作者单位】：郑州大学信息工程学院;
【基金】：国家自然科学基金(61170223,61502434)资助
【分类号】：TP311.13

【相似文献】