Loading...
Citations
Altmetric:
Abstract
A dynamic network is a network whose structure changes because of the emergence and disappearance of node or edges. It can be used to study complex systems where individuals in a system are represented as nodes and their relations/interactions are represented as edges. Studying dynamic network structures helps to better understand changes in relationships. Considerable work has been conducted on learning network structure. However, due to the complexity of dynamic networks, there is considerable room for improvement to obtain better analysis results. This thesis studies different aspects of characteristic and dynamics of a network, focusing on their application in link prediction between nodes, temporal community detection and network representation. In the first part of the thesis, we study bipartite networks constructed from online dating website data where nodes represent users and edges represent user interaction such as the exchange of invitation messages. We first formulate the prediction of future interaction between users as a link prediction problem, then propose a latent Dirichlet allocation (LDA) method to model user preferences and predict edges such that a recommendation system is built to recommend potential partners for a user. We find that user preferences changes over time and our method can adapt to these changes and outperforms baseline methods. In the second part of the thesis, we consider more general dynamic networks and model the changes in similarities between nodes over time. We present network generative models using these similarities to detect communities and their lifetime. We present a low-rank tensor decomposition technique to learn the generative models. We show that our model is robust to the change in time granularity of network during analysis and has the best performance compared to baseline methods. Finally, the last contribution of the thesis focuses on network graphlets, non-isomorphic subgraphs that represent node connection patterns in a network. We compute the significance of the graphlets by comparing the graphlet counts in an empirical network to random graphs and use this significance as feature representations for networks to analyze and characterize directed networks. Experiments show that our approach for network representation can significantly improve the accuracy on the-state-of-the-arts in network classification problem such as identifying departments in an email-exchange network or detect mobile users given their app-switching behavior represented as temporal networks.
Type
dissertation
Date
2019-02