Courses in winter term 2022 / Seminar Master-Seminar: Data Analytics 1
Instructor: Daniela Thyssens
Abstract
This seminar is directed to first term students in the Data Analytics Master Program, but students from other programs (B.Sc./M.Sc. IMIT, B.Sc./M.Sc. Wirtschaftsinformatik) are also eligible and will be admitted based on availability, please visit the corresponding LSF page for further information in this regard.
The topic for this semester's Data Analytics I Seminar is titled "Learning on Graphs".
Numerous important real world problems can be framed as learning from and on graphs. Be it social networks, biological protein-protein networks or transportation networks, any machine learning task related to these domains, would need to rely on how well graphical data can be represented, propagated and aggregated. For example, given a graphical representation of a social network where one might want to predict whether there is a link between two people, one would want to include pairwise properties between nodes, such as the number of common friends. Even though, the identification of most task-related features is quite intuitive, the technical implementation of including (encoding) this high-dimensional, non-Euclidean structural information into a vector representation remains challenging. The field of graph representation learning considers this challenge as an actual machine learning task of its own, where the goal is to learn a mapping of structural graph information on to a low-dimensional embedding space, such that the geometric relationships in the embedding space mirrors the graph structure.
Besides learning the representation of graph structures itself, many machine learning models rely on the embedded representation as a feature input to other machine learning tasks, where the representations of a graph might not even be static.
Given the ubiquitous nature of graph data and the multitude of real world applications, machine learning models that are based on graphs need to feature a pool of customizable techniques and architectural blocks to incorporate special graph structures into neural networks.
Addressing the level of first term MSc. Data Analytics students, the seminar covers key principles and building blocks that form modern Graph-based Neural Network Architectures. Furthermore, the seminar discusses classical works that revolutionized the way graphical information is encoded and covers concurrent literature picks that constitutes today's state-of-the-art in graph representation learning as well as some of its applications for machine learning downstream tasks such as forecasting or in the realm of Learning to Optimize.
Some general introductory and overview Literature to start with:
Bacciu, D., Errica, F., Micheli, A., & Podda, M. (2020). A gentle introduction to deep learning for graphs. Neural Networks, 129, 203-221.
Chami, I., Abu-El-Haija, S., Perozzi, B., Re;, C., & Murphy, K. (2022). Machine Learning on Graphs: A Model and Comprehensive Taxonomy. Journal of Machine Learning Research, 23(89), 1-64.
