In this book we have introduced several new techniques for performing web content mining tasks when utilizing more descriptive graphs in lieu of the usual case of vector representations. Our first contribution is presenting a number of ways by which web document content can be modeled as graphs. These graph representations retain information that is usually lost when using a vector model, such as term order and document section information. We demonstrated how with careful selection of a graph representation, namely a representation with unique node labels, we can perform the graph similarity task in O( n ²) time ( n being the number of nodes). In general, graph similarity using maximum common subgraph is an NP-complete problem, so this is an important result that allows us to forgo sub-optimal approximation approaches and find the exact solution in polynomial time.

Another contribution of this work is far more wide reaching: we extended classical, well-known machine learning techniques, such as the k-means clustering algorithm and k-nearest neighbors classification algorithm, to allow them to work directly with graphs as data items, instead of more limited vectors. This is a major contribution because: 1. it allows for complex, structured data, such as web documents, to be represented by a more robust model that has the potential to retain information that is usually discarded when using a vector representation and 2. we can use many existing, proven machine learning algorithms with these graph representations without having to create new,...