Walking Forward and Backward : towards Graph - Based Searching

Graphs are powerful vehicles to represent data objects that are interconnecting or inter-acting with each other. We explore random walks on various kinds of graph to addressdifferent searching and mining scenarios. This dissertation focuses on two symmetric formsof random walk called the forward walk and backward walk, which can be applied to enrichthree key tasks in searching and mining, namely, extraction, ranking and classification, innovel ways. More specifically, we enhance extraction with the metrics of probabilistic pre-cision and recall, ranking with the senses of importance and specificity, and classificationwith heterogeneous contexts in terms of relationship type and confidence level. We furtherstudy the underpinning principle of random walks on a graph, which is often known as thesmoothness assumption. We argue that smoothness is a pointwise property and requiresprobabilistic modeling. Thus, we propose a new framework to define pointwise smoothnessprobabilistically on a graph, which unifies two different “modes” of smoothness correspondingto the forward and backward random walks, respectively. Finally, our graph-based randomwalk solutions have consistently demonstrated promising empirical results for a wide rangeof searching and mining applications.