Graph mining on static, multiplex and attributed networks
Graph structured data is pervasive and generated by online human interactions at an unprece- dented velocity. Sophisticated features encoded by relational data, such as structure and relative proximity have unparalleled expressiveness to describe the complex human systems that gen- erated the data. Relational data poses challenges for information extraction and knowledge discovery due to its web scale size, extreme sparsity, multimodality, the presence of spatial autocorrelation and heterogeneity. The main goal of this thesis is to introduce principled and applicable graph mining algorithms and software packages which can tackle these challenges. Novel techniques introduced in this thesis were influenced by ideas in unsupervised machine learning, statistics and game theory. These algorithms exploit structural and proximity-based context by decomposing feature matrices, defining novel classes of characteristic functions and differentiable parametric statistical models. Proposed methods are designed to be scalable and memory efficient. The utility of machine learning methods and software packages introduced in the thesis is established by mathematical proofs and experiments that use real world networks and synthetic graphs.