In bibliometrics, and citation analysis studies in particular, many attempts have been made to analyze the relationship among scientific papers, authors and journals. Recently, these research results are found to be effective for analyzing the link structure of www pages as well. We explore the application of kernel methods to citation analysis. Kernel-based machine learning methods are getting popular due to its power and its ability to deal with (semi-)structured data such as strings and graphs. We show that a family of kernels on graphs provides a unified perspective on the three bibliometric measures that have been discussed independently: relatedness between documents, overall importance of individual documents, and importance of documents relative to one or more (root) documents (relative importance). The framework provided by the kernels establishes relative importance as an intermediate between relatedness and overall importance, in which the degree of `relativity,' or the bias between relatedness and importance, is naturally controlled by a parameter.