摘要:
近年来,具有典型多源异构特性的跨媒体数据的快速涌现给数据分析带来巨大挑战.然而,绝大多数现有跨媒体数据分析方法仅依赖模态间的共享信息发掘跨媒体数据中蕴含的模式结构,忽略各模态自身的重要信息.针对此问题,提出共享和私有信息最大化(share and private information maximization)的跨媒体聚类算法,通过兼顾跨媒体数据的共享和私有信息,以求得更加合理的聚类模式.首先,提出2种跨媒体数据的共享信息构建模型：1)混合单词模型,该模型将各模态的底层特征转换为统一的词频向量表示,然后使用一种新的自凝聚信息最大化方法自底向上地构建多模态的混合单词空间,最大化地保持各模态底层特征的统计相似性;2)聚类集成模型,构建各模态自身的聚类划分,通过互信息度量各模态聚类划分间的信息量,抽取各模态的高层聚类划分之间的相关性.其次,提出基于信息论的目标函数,将跨媒体数据的共享和私有信息融合在同一目标函数中,在抽取聚类模式结构的过程中兼顾跨媒体数据的共享和私有信息.最后,采用顺序“抽取-合并”过程优化SPIM算法的目标函数,保证其收敛到局部最优解.在6种跨媒体数据上的实验结果表明SPIM算法的优越性.
Abstract:
Recently, the rapid emergence of cross media data with typical multi-source and heterogeneous characteristic brings great challenges to the traditional data analysis approaches. However, the most of existing approaches for cross media data heavily rely on the shared latent feature space to construct the relationships between multiple modalities, while ignoring the private information hidden in each modality. Aiming at this problem, this paper proposes a novel share and private information maximization (SPIM) algorithm for cross media data clustering, which leverages the shared and private information into the clustering process. Firstly, we present two shared information construction models: 1) Hybrid words (H-words) model. In this model, the low-level features in each modality are transformed into words or visual words co-occurrence vector, then a novel agglomerative information maximization is presented to build the hybrid word space for all modalities, which ensures the statistical correlation between the low-level features of multiple modalities. 2) Clustering ensemble (CE) model. This model adopts the mutual information to measure the similarity between the clustering partitions of different modalities, which ensures the semantic correlation of the high-level clustering partitions. Secondly, SPIM algorithm integrates the shared information of multiple modalities and the private information of individual modalities into a unified objective function. Finally, the optimization of SPIM algorithm is performed by a sequential “draw-and-merge” procedure, which guarantees the function converge to a local maximum. The experimental results on 6 cross media datasets show that the proposed approach compares favorably with the existing state-of-the-art cross-media clustering methods.
百度学术
下载:
