Exploiting the Hierarchical Structure for Link Analysis

1
Exploiting the Hierarchical Structure for Link
Analysis

• Gui-Rong Xue, Qiang Yang, Hua-Jun Zeng, Yong Yu,
  Zheng Chen
• Presented by Xiaoguang Qi

2
Introduction

• Existing link analysis algorithms often suffer
  from two problems
• Sparsity of link graph
• Biased-ranking of newly-emerging pages
• Incorporate the inherent hierarchical structure
  of the web into link analysis to deal with these
  problems

3
Sketch of Hierarchical Ranking Algorithm

1. Web pages are aggregated based on their
   hierarchical structure at directory, host or
   domain level
2. Link analysis if performed on the aggregated
   graph
3. The importance of each node on the aggregated
   graph is distributed to individual pages belong
   to the node

4
Two-Layer Hierarchical Graph

• Upper-layer graph
• Partition the page set on a certain level
• One supernode for each partition
• Edges between supernodes are weighted
• Weight (Si?Sj) links from pages in Si to
  pages in Sj
• Lower-layer graph
• All the pages within a supernode are organized in
  a hierarchical structure based on the URL
  relationship

5
Hierarchical Random Walk Model

• Surf on the lower-layer graph
• Go to another page within current supernode
• Surf on the upper-layer graph
• Follow a link originated from current supernode
• Jump to a random supernode

6
Calculating Supernode Importance

• Supernode importance
• In matrix form

7
Calculating Page Importance

• Constructing weighted tree structure
• Calculating page importance by DHC

8
Parameter Tuning

• Aggregation level
• Host level aggregation is the best choice
• Parameter tuning
• ?0.6
• a0.6
• ß0.4
• ?0.8

9
Experimental Results

• Hierarchical ranking algorithm consistently
  outperforms other well-known ranking algorithms
• BM2500, BlockRank, PageRank, LayerRank,
  WeightedRank, HostRank
• Ranking on sparse data
• Effectively alleviate the sparse link problem

10
Experimental Results (Cont.)

• Ranking of new pages
• Aim to assign reasonable rank to newly-emerging
  web pages
• Test in an analogous way
• Test set 10,000 pages randomly selected with
  different rank values
• Remove 90 of their incoming links
• Perform algorithms on the modified graph