Graph Clustering for Keyword Search

Keyword search on data represented as graphs, is receiving lot of attention in recent years. Initial versions of keyword search systems assumed that the graph is memory resident. However, there are applications where the graph can be much larger than the available memory. This led to the development of search algorithms which search on a smaller memory resident summary graph (supernode graph), and fetch parts of the original graph from the disk, only when required. In this scenario, good clustering of nodes into supernodes, when constructing the summary graph, is a key to efficient search. In this paper, we address the issue of graph clustering for keyword search, using a technique based on random walks. We propose an algorithm, which we call Modified Nibble clustering algorithm, that improves upon the Nibble algorithm proposed earlier. We outline several policies that can improve its performance. Then, we compare our algorithm with two graph clustering algorithms proposed earlier, EBFS and kMetis. Our performance metrics include edge compression, keyword search performance, and the time and space overheads for clustering. Our results show that Modified Nibble outperforms EBFS uniformly, and outperforms kMetis in some settings. Further, the memory requirements of our algorithm are much lower than that of kMetis, making it practical even with a very large number of nodes, unlike kMetis.