In this paper we describe a classification method that allows the use of graph-based representations of data instead of traditional vector-based representations. We compare the vector approach combined with the k-Nearest Neighbor (k-NN) algorithm to the graph-matching approach when classifying three different web document collections, using the leave-one-out approach for measuring classification accuracy. We also compare the performance of different graph distance measures as well as various document representations that utilize graphs. The results show the graph-based approach can outperform traditional vector-based methods in terms of accuracy, dimensionality and execution time.
In this dissertation we introduce several novel techniques for performing data mining on web documents which utilize graph representations of document content. Graphs are more robust than typical vector representations as they can model structural information that is usually lost when converting the original web document content to a vector representation. For example, we can capture information such as the location, order and proximity of term occurrence, which is discarded under the standard document vector representation models. Many machine learning methods rely on distance computations, centroid calculations, and other numerical techniques. Thus many of these methods have not been applied to data represented by graphs since no suitable graphtheoretical concepts were previously available. We introduce the novel Graph Hierarchy Construction Algorithm (GHCA), which performs topic-oriented hierarchical clustering of web search results modeled using graphs. The system we created around this new algorithm and its prior version is compared with similar web search clustering systems to gauge its usefulness. An important advantage of this approach over conventional web search systems is that the results are better organized and more easily browsed by users. Next we present extensions to classical machine learning algorithms, such as the k-means clustering algorithm and the k-Nearest Neighbors classification algorithm, which allows the use of graphs as fundamental data items instead of vectors. We perform experiments comparing the performance of the new graph-based methods to the traditional vector-based methods for three web document collections. Our experimental results show an improvement for the graph approaches over the vector approaches for both clustering and classification of web documents. An important advantage of the graph representations we propose is that they allow the computation of graph similarity in polynomial time; usually the determination of graph similarity with the techniques we use is an NP-Complete problem. In fact, there are some cases where the execution time of the graph-oriented approach was faster than the vector approaches.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.