TIGRA

Many complex systems of high technological, social and scientific importance can be represented by dynamically evolving graphs. Consequently, in the last two decades we are witnessing a huge increase of valuable data structured in the form of dynamic graphs. To apply well-established and proven machine learning algorithms to such data it is necessary to form dynamic graph embeddings in Euclidean spaces. However, current research efforts have been mostly focused on embedding static graphs, while methods for dynamic graphs have not yet been extensively explored and evaluated. Existing dynamic graph embedding algorithms provide various mechanisms to incorporate temporality into the main principles of static graph embedding learning, but they do not take higherorder structural properties of graphs into account. This project will investigate the impact of hubness (high connectedness) and LID (local intrinsic dimensionality) to various quality aspects of dynamic graph embeddings. Afterwards, we will design novel hub-aware and LID-aware dynamic graph embedding methods. Additionally, we will propose the first algorithms that are able to embed graphs with time-series attributes into Euclidean spaces. We expect the project to produce novel dynamic graph embedding methods substantially more accurate than the state-of-the-art in three aspects: (1) preserving higher-order graph structural properties, (2) evolutionary stability, and (3) the accuracy of derived machine learning models for classification, clustering, event prediction, anomaly detection and diffusion prediction. By extending the graph embedding problem to graphs with time-series attributes we will propose the first dynamic graph embedding methods tackling multi-scale evolution in graph data. Consequently, the proposed methods will provide more accurate exploratory and predictive data analytics in complex dynamical systems such as e-commerce platforms, online social media, power-grids and epidemic spreading.