Computer Science Theses and Dissertations
Permanent URI for this collectionhttp://hdl.handle.net/1903/2756
Browse
2 results
Search Results
Item Computationally Comparing Biological Networks and Reconstructing Their Evolution(2012) Patro, Robert; Kingsford, Carleton L; Computer Science; Digital Repository at the University of Maryland; University of Maryland (College Park, Md.)Biological networks, such as protein-protein interaction, regulatory, or metabolic networks, provide information about biological function, beyond what can be gleaned from sequence alone. Unfortunately, most computational problems associated with these networks are NP-hard. In this dissertation, we develop algorithms to tackle numerous fundamental problems in the study of biological networks. First, we present a system for classifying the binding affinity of peptides to a diverse array of immunoglobulin antibodies. Computational approaches to this problem are integral to virtual screening and modern drug discovery. Our system is based on an ensemble of support vector machines and exhibits state-of-the-art performance. It placed 1st in the 2010 DREAM5 competition. Second, we investigate the problem of biological network alignment. Aligning the biological networks of different species allows for the discovery of shared structures and conserved pathways. We introduce an original procedure for network alignment based on a novel topological node signature. The pairwise global alignments of biological networks produced by our procedure, when evaluated under multiple metrics, are both more accurate and more robust to noise than those of previous work. Next, we explore the problem of ancestral network reconstruction. Knowing the state of ancestral networks allows us to examine how biological pathways have evolved, and how pathways in extant species have diverged from that of their common ancestor. We describe a novel framework for representing the evolutionary histories of biological networks and present efficient algorithms for reconstructing either a single parsimonious evolutionary history, or an ensemble of near-optimal histories. Under multiple models of network evolution, our approaches are effective at inferring the ancestral network interactions. Additionally, the ensemble approach is robust to noisy input, and can be used to impute missing interactions in experimental data. Finally, we introduce a framework, GrowCode, for learning network growth models. While previous work focuses on developing growth models manually, or on procedures for learning parameters for existing models, GrowCode learns fundamentally new growth models that match target networks in a flexible and user-defined way. We show that models learned by GrowCode produce networks whose target properties match those of real-world networks more closely than existing models.Item Prediction, evolution and privacy in social and affiliation networks(2011) Zheleva, Elena; Getoor, Lise; Computer Science; Digital Repository at the University of Maryland; University of Maryland (College Park, Md.)In the last few years, there has been a growing interest in studying online social and affiliation networks, leading to a new category of inference problems that consider the actor characteristics and their social environments. These problems have a variety of applications, from creating more effective marketing campaigns to designing better personalized services. Predictive statistical models allow learning hidden information automatically in these networks but also bring many privacy concerns. Three of the main challenges that I address in my thesis are understanding 1) how the complex observed and unobserved relationships among actors can help in building better behavior models, and in designing more accurate predictive algorithms, 2) what are the processes that drive the network growth and link formation, and 3) what are the implications of predictive algorithms to the privacy of users who share content online. The majority of previous work in prediction, evolution and privacy in online social networks has concentrated on the single-mode networks which form around user-user links, such as friendship and email communication. However, single-mode networks often co-exist with two-mode affiliation networks in which users are linked to other entities, such as social groups, online content and events. We study the interplay between these two types of networks and show that analyzing these higher-order interactions can reveal dependencies that are difficult to extract from the pair-wise interactions alone. In particular, we present our contributions to the challenging problems of collective classification, link prediction, network evolution, anonymization and preserving privacy in social and affiliation networks. We evaluate our models on real-world data sets from well-known online social networks, such as Flickr, Facebook, Dogster and LiveJournal.