Open topics for theses and practical courses

Graph Neural Networks (GNNs) have significantly enhanced the capability of recommendation systems in e-commerce by leveraging collaborative filtering and session-based methodologies. These advancements have demonstrated improved recommendation accuracy by effectively capturing complex user-item interactions within graph structures. However, the deployment of GNN-based systems at scale encounters substantial challenges, primarily due to their computational intensity and reliance on GPU hardware, which becomes a bottleneck in scenarios dealing with millions of requests per second. This thesis aims to address the scalability issues inherent in GNNs for e-commerce recommendation systems by exploring knowledge distillation techniques. The objective is to transfer knowledge from complex GNN models to simpler, more computationally efficient representations that can be easily managed and scaled on CPU servers and database environments. This research will involve a comprehensive evaluation of existing knowledge distillation methods applied to GNNs, focusing on their effectiveness in preserving the recommendation quality while significantly reducing computational requirements. A novel methodology for fair comparison of these techniques shall be developed, taking into account factors such as recommendation accuracy, computational efficiency, and scalability. The outcome of this thesis is expected to provide a systematic approach for making GNN-based recommendation systems more practical for large-scale e-commerce applications, thereby bridging the gap between advanced machine learning models and their real-world applicability in resource-constrained environments.

Requirements: Python, interest in recommendation systems, interest in knowledge distillation

Contact: Lorenz Kummer, Nils Kriege

Human cells need to precisely regulate the amount of proteins produced from each gene. Many of the regulatory mechanisms are governed by specific properties of the messenger RNA (mRNA). However, only a few of these properties have been studied in detail. This project employs deep learning to identify new and relevant RNA features that impact translation and stability. Deep learning models will be trained to predict the lifetime of an mRNA and its protein expression from mRNA sequence and structure information. The models will then be analyzed using techniques from the field of explainable AI to extract which mRNA features were crucial for the prediction.

Students working on this project should be knowledgeable in machine learning, have solid programming skills, have a desire to work on interpretable deep learning, and be keen to use machine learning to advance science and the discovery of biological knowledge.

Contact: Sebastian Tschiatschek

Inverse reinforcement learning is about identifying the reward function optimized by an agent (e.g., an AI system or a human) through its behavior. This is for instance important to enable the transfer of a reinforcement learning agent’s capabilities to new environments and value alignment. Most commonly, inverse reinforcement learning techniques assume that the agent demonstrating optimized behavior and the agent learning from these demonstrations about the reward function have the same capabilities in terms of actions and observations. This assumption is however unrealistic and you will challenge it in this project.

This project considers inverse reinforcement learning in settings in which there is some form of embodiment mismatch between the expert demonstrator and the learning agent. The scope of the project is to study how existing algorithms perform in this setting and propose modifications to existing algorithms to achieve better performance, e.g., in the form of demonstrations tailored to the learner, learning about the mismatch, or similar.

Students wanting to work on this topic are expected to have a solid basic understanding of machine learning techniques, a solid knowledge of Python, and a basic knowledge of deep learning libraries (PyTorch or TensorFlow).

Contact: Sebastian Tschiatschek

While reinforcement learning has shown impressive performance in various applications, the learned agent’s behavior often lacks interpretability and planning is computationally demanding. One avenue to approach these challenges is learning abstractions, i.e., aggregating different states of an environment into a meta-state such that te size of the resulting meta-state space is reduced. Such abstractions are most useful if they satisfy certain properties, e.g., the resulting model is Markovian or allows for a causal interpretation.

In this project you will implement and compare existing abstraction techniques for reinforcement learning and extend a selected one to allow for hierarchical abstraction to enable interpretation and planning at different levels of granularity while ensuring the aforementioned properties.

Students wanting to work on this topic are expected to have a solid basic understanding of machine learning techniques, a solid knowledge of Python, and a basic knowledge of deep learning libraries (PyTorch or TensorFlow).

Contact: Sebastian Tschiatschek

Explainable Artificial Intelligence (XAI) focuses on providing explanations for domain experts and data scientists. However, 'lay people', i.e., people without or only with limited technical knowledge, are often supported by AI systems in public institutions in their work, e.g., to make decisions about loan applications. Thus it is important that these lay people can understand the AI systems to a certain extent and understand the impact of these systems on the decisions they make.

In this project, you will study how to best make AI systems comprehensible to lay people in sequential decision-making settings, in particular in settings motivated by the usage of AI systems in public institutions. It will involve developing AI models, interpretability techniques, and conducting user studies with lay people.

Contact: Sebastian Tschiatschek

Learning problems for molecular data are often imbalanced, i.e., there is a small subset of molecules having a desired property while most molecules do not have this property. Typically, the minority class of special interest in the considered application. However, most machine-learning algorithms do not work well in such settings. In this project, you will explore and compare several general-purpose methods for imbalanced data and develop specific data augmentation methods for molecular data.

Contact: Nils Kriege

The data regarded in this thesis is from the field of Molecular Dynamics (MD) and describes the positions of a protein's atoms over time. Clustering timesteps allows to detect different states of a protein, which is important for understanding the protein folding process. E.g., these states can serve for a Markov State Model of the protein's trajectories.

In this work, we compare results based on the usually applied molecular descriptors with 3d-Shape descriptors that were recently developed for general point clouds.

Contact: Anna Beer

Early graph neural networks (GNNs) were motivated by circular chemical fingerprints generalizing their concept by introducing learnable parameters. However, recent studies show that current state-of-the-art GNN architectures are still outperformed in molecular property prediction by concrete implementations of circular fingerprints combined with a multilayer perceptron. The task of this project is to explore the reasons for this through a detailed analysis of both methods. You will investigate a concrete implementation of a chemical fingerprint and develop a new neural version starting by reproducing the fingerprint and adding neural modules stepwise. Different stages should be investigated in an ablation study.

Contact: Nils Kriege

Explainable Artificial Intelligence (XAI) focuses on providing explanations for domain experts and data scientist. However, 'lay people', i.e., people without or only with limited technical knowledge, are often subject to the consequences of deploying algorithmic decision-making systems in public institutions, such as the COMPAS recidivism model. Finding new ways to render machine learning models comprehensible for lay people is therefore a current and highly relevant challenge. This project will focus on approaches for building machine learning models that are in themselves interpretable and explainable for lay people, soliciting slight or no external explanations. It will involve both programming and the conducting of user studies with lay people.

Contact: Timothée Schmude, Sebastian Tschiatschek

Many research question are not only statistical but rather causal in nature. Two variables may be statistical dependent on each other, but do not not cause each other. For example, the relative number of Nobel laureates for a country strongly correlates with its annual chocolate consumption. However, increasing the national chocolate consumption, does obviously not produce more outstanding scientists. It is a key challenge for machine learning to move from learning purely statistical relations to causal ones. In the absence of randomized trials, causal structure learning focuses on identifying which variables in a given dataset are directly causally related and how these direct causal relations form a structure over which also indirect causal relations exist. These dependencies can be encoded by a Directed Acyclic Graph (DAG) and the first step in performing causal inference is to determine such underlying causal graph.

In this project you will make yourself familiar with causal structure learning in machine learning. You will benchmark recent algorithms to find underlying causal relations in datasets and apply them to questionnaire data from the social and political sciences.

Students working on this project need basic background knowledge in machine learning or probability, good programming skills in Python, and the desire to implement and test causal discovery algorithms.

Contact: Simon Rittel, Sebastian Tschiatschek

Auxiliary prediction tasks have been the source of many recent performance improvements for reinforcement learning agents, particularly when working with image inputs. These tasks are usually provided as expert knowledge by the algorithm designer and force the agent to pay attention to various aspects of the environment which might be helpful to solving the underlying task. In this project you will investigate whether such prediction tasks can also be learned by the agent itself. For this, you will first visualize what an RL agent actually "attends to" through investigating heatmaps of their first network layer. Then you will devise a method allowing the agent itself to decide which pixels of the state it wants to learn to reconstruct.

Required skills: Good programming skills, basic background in deep learning

Contact: Timo Klein, Sebastian Tschiatschek

The graph edit distance (GED) quantifies the dissimilarity of two graphs as the minimum cost of a sequence of edit operations turning one graph into the other. The problem complexity depends on the choice of the edit operation costs and the properties of the considered graphs. It has been shown that the classical NP-hard subgraph isomorphism problem (SI) and the maximum common subgraph problem (MCS) can be reduced to computing the GED. As a consequence, computing the GED also is NP-hard. However, for SI and MCS a fine distinction of the complexity depending on parameters such as treewidth, degree, and their combination is known. The goal of this project is to investigate the complexity of computing the graph edit distance in restricted graph classes identifying polynomial-time solvable and NP-hard cases.

Requirements: Strong interest in graph algorithms and theoretical computer science

Contact: Nils Kriege

In this project/thesis you will study the potential of using reinforcement learning (RL) for improving mental health treatments. In particular, you will investigate the potential of RL algorithms for optimizing exposure therapy protocols, which are usually used for the treatment of anxiety disorders, such as phobias.

Specifically, you will conduct simulation studies in which a computational associative learning model (such as the latent cause model) will be a stand-in for the patients, and where an RL algorithms will make modifications to the exposure protocol with the goal of robustly extinguishing negative associations, which play a crucial role in anxiety disorders. The investigation can start with standard model-free RL approaches, but is then planned to proceed with model-based augmentations, such as equipping the RL algorithm with "Theory-of-Mind"-type capabilities.

Students working on this project need basic background knowledge in machine learning, solid programming skills, a desire to work on reinforcement learning, and willingness to learn about models of human cognition.

Contact: Sebastian Tschiatschek

Uncertain graphs represent an important data model for real-world networks, where one can assign an independent probability of existence on every edge. Uncertainty, represented by the edge probability, may arise due to several reasons, e.g., errors in measurements, data integration from inconsistent and ambiguous sources, lack of precise information needs, inference and prediction models, and explicit manipulation for privacy purposes. Due to their rich expressiveness and utility in a wide range of applications, uncertain graphs have prompted a great deal of research by the data mining research community.

This work will mainly focus on designing efficient algorithms for selecting and modifying the probability of a given number of edges in an uncertain graphs. The developed algorithm will be evaluated against other existing methods. An experimental study will be conducted on real-world datasets. Students working on this project need basic knowledge in graph theory and algorithms, and solid programming skills in Python, C/C++ or Java.

Contact: Yllka Velaj, Claudia Plant

For two given graphs, G and H, the maximum common subgraph problem (MCS) asks for the largest graph contained in both G and H. An important application occurs in cheminformatics, where the similarity of molecular graphs needs to be quantified. Unfortunately, MCS is NP-hard unless additional constraints regarding the properties of G, H, and the subgraph are enforced. A polynomial-time solvable case requires that all these graphs are trees. Known negative complexity results for classes of tree-like graphs set narrow boundaries for generalization.

The project aims to investigate particular tree-like graph classes to design, implement, and experimentally evaluate new polynomial-time algorithms. New NP-hardness proofs for specific cases should refine the complexity status further.

Requirements: Strong interest in algorithms and graph theory

Contact: Nils Kriege

A graph isomorphism is a mapping between the nodes of two graphs that preserves the edges. An automorphism is an isomorphism of a graph to itself. The (non-trivial) automorphisms intuitively represent the symmetries of a graph. Two vertices are considered equivalent if there is an automorphism that maps one vertex to the other. The corresponding equivalence classes are referred to as orbit partition. The orbit partition is of interest for symmetry breaking techniques for various combinatorial problems. Identifying and pruning symmetric branches in the search tree based on the symmetries of the input instance can lead to drastic speed-ups. This project aims to develop dynamic algorithms for maintaining the orbit partition when the input graph changes. The focus of the project can be on the development and theoretical analysis of algorithms or their implementation and experimental evaluation.

Requirements: Strong interest in algorithms

Contact: Nils Kriege, Christian Schulz, Kathrin Hanauer

In this project you will study novel approaches for learning interpretable and explainable deep learning models with the goal of supporting users with different skill levels and knowledge. The models will be evaluated on applications with large societal impact, e.g., labour market data.

Contact: Sebastian Tschiatschek

In many areas of our life questionnaires/a collection of measurements are used to gather information for decision making, e.g., questionnaires are used to measure the success of marketing campaigns or customer satisfaction, or medical tests are conducted to decide on the best treatment for a patient. Common to those applications is that many questions need to be answered or many tests must be conducted. Often this is not because all the information is needed for deriving a decision but rather because the course of action is standardized (e.g., a survey always consists of the same questions).

In this project you will evaluate and develop methods for adaptive information acquisition that acquire the information necessary for decision making sequentially and adaptively, e.g., not all questions in a questionnaire are asked in all cases and the order of the questions can be changed. You will consider and evaluate recent machine learning models for this problem but also investigate how these methods can be extended to account for dependencies in the information that can be required, e.g., dependencies that only allow for a specific order of some questions of a questionnaire, the influence of how questions are asked on the answers, etc.

Contact: Sebastian Tschiatschek

Clustering is the task of finding groups in a data set. As an unsupervised learning approach it is important for exploratory data analysis, e.g., for finding commonalities between patients suffering from some disease. Often such data is high-dimensional and contains complicated non-linear relationships, which makes it difficult to apply standard clustering approaches. Deep clustering algorithms have been proposed recently to approach this kind of setting by combining clustering and deep learning in a common framework. An open problem that has not been sufficiently addressed in recent deep clustering research is incomplete and heterogeneous (e.g. survey data, blood tests, MRI images, etc.) data. Incomplete data is often an issue in practice, especially for health data where not everything about a patient’s history is known.

In this project you will study the clustering of incomplete data with heterogeneous data types with a particular focus on missingness and on determining which information is relevant for cluster assignments. You will develop an algorithm based on previous work and evaluate it on synthetic and real-world data. Some research questions of interest would be for instance: How do cluster assignments change when given new information? Or, given an uncertain cluster assignment (e.g. sick or healthy patient) which information would we need to query to make a decision?

Contact: Sebastian Tschiatschek

Enabling intelligent agents to collaborate effectively despite mismatch in perception or capabilities.

Contact: Sebastian Tschiatschek

Enabling multiple interacting agents to convey their knowledge and skills through teaching.

Contact: Sebastian Tschiatschek

Learning abstractions of large state and action spaces in order to facilitate more efficient planning and exploration.

Contact: Sebastian Tschiatschek

Bivariate causal measures: Experiments with existing causal methods in python on synthetic and real data.

Contact: Katerina Schindlerova

Please see [Moodle] for a detailed description of this and other Natural Language Processing topics.

Contact: Benjamin Roth

Please see [Moodle] for a detailed description of this and other Natural Language Processing topics.

Contact: Benjamin Roth

Please see [Moodle] for a detailed description of this and other Natural Language Processing topics.

Contact: Benjamin Roth

Please see [Moodle] for a detailed description of this and other Natural Language Processing topics.

Contact: Benjamin Roth

Please see [Moodle] for a detailed description of this and other Natural Language Processing topics.

Contact: Benjamin Roth

Please see [Moodle] for a detailed description of this and other Natural Language Processing topics.

Contact: Benjamin Roth

Please see [Moodle] for a detailed description of this and other Natural Language Processing topics.

Contact: Benjamin Roth

Please see [Moodle] for a detailed description of this and other Natural Language Processing topics.

Contact: Benjamin Roth

Reinforcement learning has shown remarkable successes in playing board games like Go but also in playing computer games like DOTA. While these results are impressive, the learned strategies (policies) are typically not directly interpretable by humans. However, interpretability is important for instance to enable better collaboration of humans with AI agents or to assess safety of the learned behavior.

In this project you will make such learned policies easier to understand by humans by incorporating constraints on their structure. You will develop and implement a novel policy training algorithm and evaluate it on simple computer games.

Students wanting to work on this topic are expected to have solid coding skills in Python, basic knowledge in PyTorch or TensorFlow, and be curious to learn about reinforcement learning and its applications.

Contact: Sebastian Tschiatschek

Reinforcement learning has shown remarkable successes in application domains like game play and robotics but suffers from large sample complexity, limiting its broader usage. A promising approach for reducing the sample complexity is to learn skills that can be reused when solving a series of related tasks. In the context of robotics, such skills could for instance correspond to lifting a leg or grepping an object.

In this project you will develop and implement a learning algorithm algorithm for such composable policies based on recently proposed Successor Options. You will evaluate this algorithm on several benchmark tasks and compare it to other state of the art algorithms.

Students wanting to work on this topic are expected to have solid coding skills in Python, basic knowledge in PyTorch or TensorFlow, and be curious to learn about reinforcement learning and its applications.

Contact: Sebastian Tschiatschek

Reinforcement learning has shown remarkable successes in application domains like game play and robotics. These successes where achieved by performing large numbers of interactions of a learning agent with the environment. In these interactions, the agent constantly receives information about rewarding behavior. However, in many applications involving human users (e.g., personal assistants), the reward information is explicitly or implicitly provided by these human users. Therefore, this information must be treated as a scarce and valuable resource, and the (cognitive) cost of providing this information should be accounted for.

In this project you will study the cost of providing different types of feedback used for reinforcement learning in user studies. To this end, you will implement a simple reinforcement learning environment for which a human user can provide reward information in various forms (e.g., comparisons, sorted lists, etc.) and compare the cost of providing this feedback as well as the usefulness of the feedback for training the reinforcement learning agent. Depending on whether you are doing a P1/P2/Bachelor Thesis/Master Thesis, the scope of the project will be adjusted.

Students wanting to work on this topic are expected to have solid coding skills in Python, basic knowledge about the development of interactive web pages, and be curious to learn about reinforcement learning and its applications.

Contact: Sebastian Tschiatschek

Recent advances in reinforcement learning have enabled super-human performance of AI agents in many applications, e.g., game play. In this project you will work on building theory and/or models for reinforcement learning from implicit and explicit feedback. In particular, you will consider the setting in which no rewards are observed during execution of the agent but only cumulative reward information is obtained at the end of the execution. As a practical example, assume playing a computer game in which you cannot observe the score while you play the game but only see your achieved score at the very end. This setting has important applications in human-in-the-loop settings in which feedback can be expensive to obtain.

Students working on this project need basic background knowledge in machine learning, solid programming skills in Python, and the desire to work on reinforcement learning.

Contact: Sebastian Tschiatschek

In this project you will work on predicting students' responses to mathematical questions using machine learning models. This is important to assess a student's skills and provide personalized educational resources. We use real-world data provided by the "Diagnostic Questions: Predicting Student Responses and Measuring Question Quality" challenge at NeurIPS'20. The challenge consists of four different tasks from which a subset can be selected for the project, depending on background knowledge and programming skills.

Students working on this project need basic background knowledge in machine learning, programming skills in Python, and the desire to develop and test machine learning models.

Contact: Sebastian Tschiatschek

Automated sequential decision making is an important application of machine learning systems in which such a system needs to select a sequence of actions step by step to optimize a reward/utility function. For instance, in autonomous driving, such a system needs to execute a sequence of steering, breaking and acceleration actions, or in a medical intensive care setting, such a system needs to execute a sequence of measurement and treatment actions.

One challenge in realizing such automated sequential decision making systems is the definition of the reward/utility function. For example, in autonomous driving it is hard to specify all the factors which define good driving behavior. In such settings, automatically inferring the reward/utility function from users’ feedback can be beneficial.

This project investigates approaches for reward/utility inference from diverse and implicit feedback, building on ideas for inverse reinforcement learning, active learning, implicit feedback, etc.

Interested students are expected to have solid mathematical and machine learning skills, and have experience in Python and deep learning (using PyTorch or TensorFlow).

Contact: Sebastian Tschiatschek

Reinforcement learning has been successfully used to solve certain challenging sequential decision making problems in recent years. The employed techniques commonly require (i) huge amounts of interactions with the envirnoment and (ii) clearly specified reward signals to work well. In many applications however, one or both of these requirements are not met. In such cases, imitation learning can be an efficient approach to sequential decision making problems: an expert demonstrates near-optimal behavior and a learning agent attempts to mimic this behavior.

This project considers imitation learning in settings in which there is some form of mismatch between the expert demonstrator and the learning agent. The scope of the project is to study how existing algorithms perform in this setting and proposes modifications to existing algorithms to achieve better performance.

Students wanting to work on this topic are expected to have a basic understanding of machine learning techniques, solid knowledge of Python and basic knowledge of deep learning libraries (PyTorch or TensorFlow).

Contact: Sebastian Tschiatschek

Many applications involve the selection of sequences of items, e.g., recommender systems and personalization in MOOCs. Common to many of these applications is that the order of the selected items is not random but that items selected early, influence which items are selected later. The problem of selecting sequences of items has been studied in various settings, including that in which dependencies between items are modeled using graphs and monotone submodular functions.

This project aims at extending these settings to cover the case of non-monotone submodular functions, by proposing new algorithms and analyzing their properties. The findings are validated by an implementation of the proposed algorithm(s) and comparison agains reasonable baselines.

Students wanting to work on this topic are expected to have solid mathematical skills, a basic understanding of machine learning techniques, solid knowledge of Python and deep learning libraries (PyTorch or TensorFlow).

Contact: Sebastian Tschiatschek

Variational Autoencoders (VAEs) are powerful deep generative models that have been successfully applied in a wide range of machine learning applications. Recently, the Partial VAE (PVAE), a variant of VAEs that can process partially observed inputs has been proposed and its effectiveness for data imputation has been demonstrated. Key to the fast training of VAEs and PVAEs is the amortized prediction of posterior distributions from observations. In PVAEs, these posterior distributions are predicted from partial observations.

This project aims at studying the consistency of these posterior distributions for different patterns of missing data. The insights are used to create/train better inference models and thereby improve the quality of PVAEs.

Students wanting to work on this topic are expected to have solid mathematical skills, a basic understanding of machine learning techniques and good programming skills in Python.

Contact: Sebastian Tschiatschek

This work will focus mainly on implementation of an algorithm for causal detection among processes as well as its testing on probability distributions with heavy tailed distributions and on climatological data provided by ZAMG (Die Zentralanstalt für Meteorologie und Geodynamik) following these distributions. More details to the interested student will be provided on request.

Contact: Katerina Schindlerova