Probabilistic Random Walk Models for Comparative Network Analysis

Abstract

Graph-based systems and data analysis methods have become critical tools in many fields as they can provide an intuitive way of representing and analyzing interactions between variables. Due to the advances in measurement techniques, a massive amount of labeled data that can be represented as nodes on a graph (or network) have been archived in databases. Additionally, novel data without label information have been gradually generated and archived. Labeling and identifying characteristics of novel data is an important first step in utilizing the valuable data in an effective and meaningful way. Comparative network analysis is an effective computational means to identify and predict the properties of the unlabeled data by comparing the similarities and differences between well-studied and less-studied networks. Comparative network analysis aims to identify the matching nodes and conserved subnetworks across multiple networks to enable a prediction of the properties of the nodes in the less-studied networks based on the properties of the matching nodes in the well-studied networks (i.e., transferring knowledge between networks).

One of the fundamental and important questions in comparative network analysis is how to accurately estimate node-to-node correspondence as it can be a critical clue in analyzing the similarities and differences between networks. Node correspondence is a comprehensive similarity that integrates various types of similarity measurements in a balanced manner. However, there are several challenges in accurately estimating the node correspondence for large-scale networks. First, the scale of the networks is a critical issue. As networks generally include a large number of nodes, we have to examine an extremely large space and it can pose a computational challenge due to the combinatorial nature of the problem. Furthermore, although there are matching nodes and conserved subnetworks in different networks, structural variations such as node insertions and deletions make it difficult to integrate a topological similarity.

In this dissertation, novel probabilistic random walk models are proposed to accurately estimate node-to-node correspondence between networks. First, we propose a context-sensitive random walk (CSRW) model. In the CSRW model, the random walker analyzes the context of the current position of the random walker and it can switch the random movement to either a simultaneous walk on both networks or an individual walk on one of the networks. The context-sensitive nature of the random walker enables the method to effectively integrate different types of similarities by dealing with structural variations. Second, we propose the CUFID (Comparative network analysis Using the steady-state network Flow to IDentify orthologous proteins) model. In the CUFID model, we construct an integrated network by inserting pseudo edges between potential matching nodes in different networks. Then, we design the random walk protocol to transit more frequently between potential matching nodes as their node similarity increases and they have more matching neighboring nodes. We apply the proposed random walk models to comparative network analysis problems: global network alignment and network querying. Through extensive performance evaluations, we demonstrate that the proposed random walk models can accurately estimate node correspondence and these can lead to improved and reliable network comparison results.