Privacy Preserving Link Analysis on Dynamic Weighted Graph


DOI: 10.1007/s10588-005-3941-2

Cite this article as:
Duan, Y., Wang, J., Kam, M. et al. Comput Math Organiz Theor (2005) 11: 141. doi:10.1007/s10588-005-3941-2


Link analysis algorithms have been used successfully on hyperlinked data to identify authoritative documents and retrieve other information. They also showed great potential in many new areas such as counterterrorism and surveillance. Emergence of new applications and changes in existing ones created new opportunities, as well as difficulties, for them: (1) In many situations where link analysis is applicable, there may not be an explicit hyperlinked structure. (2) The system can be highly dynamic, resulting in constant update to the graph. It is often too expensive to rerun the algorithm for each update. (3) The application often relies heavily on client-side logging and the information encoded in the graph can be very personal and sensitive. In this case privacy becomes a major concern. Existing link analysis algorithms, and their traditional implementations, are not adequate in face of these new challenges. In this paper we propose the use of a weighted graph to define and/or augment a link structure. We present a generalized HITS algorithm that is suitable for running in a dynamic environment. The algorithm uses the idea of “lazy update” to amortize cost across multiple updates while still providing accurate ranking to users in the mean time. We prove the convergence of the new algorithm and evaluate its benefit using the Enron email dataset. Finally we devise a distributed implementation of the algorithm that preserves user privacy thus making it socially acceptable in real-world applications.


link analysis data mining text analysis privacy HITS graph algorithms lazy update 

Copyright information

© Springer Science + Business Media, Inc. 2005

Authors and Affiliations

  • Yitao Duan
    • 1
  • Jingtao Wang
    • 1
  • Matthew Kam
    • 1
  • John Canny
    • 1
  1. 1.Computer Science DivisionUniversity of California at BerkeleyBerkeley

Personalised recommendations