Abstract
This paper presents an accurate and efficient method for extracting hierarchical structure of Web communities, i.e., Web video sets with similar topics for Web video retrieval. First, efficient canonical correlation analysis (CCA), named sub-sampled CCA, is derived to obtain link relationships that represent similarities between latent features of Web videos. Moreover, the obtained link relationships enable application of an algorithm based on recursive modularity optimization to extract hierarchical structure of Web communities. Different from previously reported methods, our method can extract the hierarchical structure for the whole target dataset since the algorithm enables recursive reduction of its processing targets. This means it becomes unnecessary to perform screening of Web videos, and we can avoid performance degradation caused by discarding relevant Web videos in the screening, which occurred in previously reported methods. Consequently, our method enables extraction of the hierarchical structure with high accuracy as well as low computational cost.
