Probabilistic and Interactive Machine Learning

The work group Probabilistic and Interactive Machine Learning is headed by Ass.-Prof. Sebastian Tschiatschek. It is part of the research group Data Mining and Machine Learning.


Our research is in developing and understanding machine learning systems to support decision making in uncertain domains. We are particularly interested in systems that learn and adapt, requiring such systems to actively acquire relevant information and reason about the value of that information, and decision making systems which face or include human users.

We focus on the development and analysis of machine learning methods in the following three broad topics:

  • Reinforcement Learning
    The basic question in reinforcement learning is how can we build effective systems that automatically learn from interacting with some unknown environment. In that regard, we are particularly interested in aspects of exploration, i.e., learning about new environments, adaptive goal-oriented active data acquisition, e.g., for learning a human user’s preferences, inverse reinforcement learning, learning from noisy and costly feedback, etc.
  • Interactive machine learning
    Many machine learning systems interact with human users or users are part of the machine learning pipeline. In interactive machine learning we focus on enabling human users and machine learning systems to work together effectively with the goal of enabling fast learning (of the users and the algorithms), users’ satisfaction, and effective and robust decision making.
  • Probabilistic Models
    Accurate probabilistic quantification of uncertainty and taking this uncertainty into account in decision making is key for dealing with the stochasticity inherent in many decision making problems. For that purpose, we develop and analyze novel probabilistic models, partly based on deep learning. In particular, we are interested in variational autoencoders and generative adversarial networks but also (more) classical probabilistic graphical models (Bayesian Networks, Sum-Product-Networks, etc.). Our research in this topic has a large focus on handling missing data, time-series modelling and using heterogeneous data.

Work Group