Polynomial Time Complexity Graph Distance Computation for Web Content Mining

Abstract

Utilizing graphs with unique node labels reduces the complexity of the maximum common subgraph problem, which is generally NP-complete, to that of a polynomial time problem. Calculating the maximum common subgraph is useful for creating a graph distance measure, since we observe that graphs become more similar (and thus have less distance) as their maximum common subgraphs become larger and vice versa. With a computationally practical method of determining distances between graphs, we are no longer limited to using simpler vector representations for machine learning applications.We can perform well-known algorithms, such as k-means clustering and k-nearest neighbors classification, directly on data represented by graphs, losing none of the inherent structural information. We demonstrate the benefits of the additional information retained in a graph-based data model for web content mining applications. We introduce several graph representations for capturing web document information and present some examples of our experimental results, which compare favorably with traditional vector methods.