Variability of Dynamic Node Embeddings

点击翻译按钮获取中文摘要

Node embedding algorithms have attracted much research interest over the last five years, with neural approaches arguably being among the most popular approaches. Many of the state-of-the-art node embedding algorithms utilize random processes that yield multiple different embeddings of the same graph even under fixed parameters. These instabilities naturally influence the outcome of any downstream task that is performed on top of node embeddings. A recent study conducted by the applicants has shown that this can particularly affect instance-level down-stream classifications. As a result, instability effects pose a sensitive issue when working with node embeddings, even more so when considering dynamically evolving graphs. In this project, we aim to develop a deep understanding of node embeddings for dynamically evolving graphs, with particular attention paid to understanding variability. In addition, we anticipate exploring efficient algorithms that are capable of embedding dynamic graphs while controlling the variability of resulting embeddings. Overall, our work contributes to the development of more principled approaches for network representation learning.