Although complex Machine Learning models may provide accurate predictions, they often suffer from being biased or difficult to explain. In this seminar, we will explore current approaches for more transparency, explainability and fairness in different areas of Machine Learning and Data Analytics.
General Information
Lecturer: Prof. Gjergji Kasneci; Organisation: Martin Pawelczyk, Tobias Leemann
The registration and organisation of the seminar will be handled via Ilias. Please sign up early as the number of places is limited.
Our first meeting will take place on Monday, 24th of October 2022 at 2:00pm in room A104 (Sand 14). There we will share further information about the content and organisation of the seminar. Usually, students are required to complete small research projects in teams up to two.
Prerequesites: There are no specific prerequisites. A solid background in Computer Science and Mathematics, statistics in particular, is beneficial.
Topics
Explainability for Online Machine Learning
- Developing Explanation Methods and Properties for Dynamic Data Streams
- Can our understanding of explainability change over time? Is post-hoc and local explainability meaningful and feasible in a data stream?
- Explainable Online Learning with Decision Rules
- Are algorithms based on decision rules more suited for explainable online learning than post-hoc attribution methods?
Conceptual Explanations
- Extending Annotated Concepts with Learned Concepts for Complete Explanation
- How can automatically annotated concepts be extended with meaningful unsupervisedly discovered concepts?
- Explaining Decisions in Terms of Connected Concepts
- Suppose we have a graph of concepts and their semantic connection (and maybe some form of representation such as a GNN). How can we explain what a GNN has learned and come up with some kind of attributions?
Counterfactual Explanations and Strategic Classification
- How can adversarial attacks impact counterfactual explanations?
- How robust are sequential counterfactual explanations?
- Can we devise strategy-robust counterfactual explanations?
- Are counterfactual explanations highlighting the most important features?
Fairness for automated decision-making systems
- Evaluating fairness for automated decision making systems (especially risk and welfare considerations).
- We want to study how the convex formulation of welfare-based measures of fairness allows us to integrate them as a constraint into any convex loss minimization pipeline.
Explainability of deep neural networks through feature attribution.
- Evaluation of local explanations