About the topic 

Representation Learning is a central paradigm in modern Machine Learning that aims to automatically learn informative, compact, and transferable representations from data. Representation learning methods discover latent structures that support downstream tasks such as classification, retrieval, recommendation, and reasoning.

In this seminar, we focus on representation learning over structured and semi-structured data, with an emphasis on graphs, knowledge graphs, and text. These data modalities are foundational in many real-world systems: graphs capture relational structure (e.g., social networks, molecules, transaction networks), knowledge graphs model semantically annotated and interconnected facts, and text encodes large amounts of world knowledge.

In this seminar, students will learn about advanced topics in learning, analyzing, and applying representations for graphs, knowledge graphs, and text.

Potential topics include:

• Fundamentals of representation learning: objectives, inductive biases, and generalization
• Graph representation learning
• Graph Neural Networks
• Self-supervised and contrastive learning on graphs and text
• Knowledge graph embeddings: translational, bilinear, and neural approaches
• Learning over knowledge graphs with GNNs
• Language models as representation learners
• How LLMs store and use knowledge: memorization, retrieval, and reasoning behavior
• Retrieval-Augmented Generation (RAG)
• Evaluation of representations: benchmarks, robustness, interpretability, bias

About the seminar

This seminar offers students an opportunity to explore research in representation learning via:

• Independent assessment of an advanced scientific theme
• Systematic presentation and discussion of scientific results
• Preparation of a term paper, including a section on related work

Previous Knowledge Expected

Basic knowledge about the following topics is highly recommended but not mandatory:

• Linear algebra and probability
• Machine learning fundamentals
• Graph theory
• Databases and knowledge representation basics