From Networks to Insights: Analyzing Graphs With and Without Machine Learning

Note: Selected to integrate a very small group of applicants to give a presentation to the faculty.

Abstract

Networks are ubiquitous, and their presence is being increasingly noticed across multiple disciplines. They represent a powerful language that distills the intricate relationships of even the most complex systems into manageable structures, allowing a systematic analysis. Despite their potential, effectively applying network science to certain fields still faces challenges. For example, the non-trivial translation process necessary to convert systems into graphs and limitations of existing techniques when faced with exotic systems. I believe, streamlining the integration of networks in more fields can be done in three ways: (1) development of more powerful, interpretable and efficient “traditional” network science algorithms, (2) direct adaptation of machine learning algorithms and techniques to function on graph data, being that at the level of knowledge extraction and process or graph translation and generation; (3) co-evolutionary refinement of traditional machine learning theory and theoretical graph-machine learning, using the fact that graphs and other geometric approaches can be an overarching system to other types of data. In my presentation, I will give a brief look into the motivations behind my choice to focus on machine learning and network science and will briefly discuss my research background, focusing on graph-related projects. I will then present some questions I would like to explore, mostly related to the three points previously highlighted while giving some hints regarding the application areas I have an interest in.