Theses
Our research group offers you the possibility to write your theses with us at any time. You can either apply for one of the suggested topics or submit your own proposal, which should be related to one of our research areas. Theses can also be conducted in cooperation with a company.
If you would like to apply for a suggested topic, please send a brief description of your interest in the topic and your last transcript of records via email to the responsible supervisor.
If you would like to propose your own topic, please send a working title, a short description of your topic and your planned approach together with your last transcript of records via email to the researcher(s) of the respective research area.
If you would like to be updated when new topics are announced, as well as to access research materials, please join our StudOn group.
Open Theses Topics
Wearable AI and XAI
Method: literature review or prototype (options to be discussed)
Language: English
Supervisor: Pavlina Kröckel
Business models of drug discovery startups
Language: English
Supervisor: Pavlina Kröckel
Health Misinformation Spread on Social Media - A benchmark of German language sources
Language: English
Required: Analytics skills – specifically text analytics; preferably also Python skills
Supervisor: Pavlina Kröckel
Data-driven decision making on the shop floor – A state-of-the-art analysis
Language: German or English
Supervisor: Markus Schamberger
Human-in-the-loop shop floor control – A state-of-the-art analysis
Language: German or English
Supervisor: Markus Schamberger
Understanding the individual value creation potential of cross-enterprise collaboration in federated networks
Motivation
Companies in almost all industries are striving to implement AI-based solutions to benefit from data-driven decision-making, optimized/automated processes, or by offering services tailored to customers’ needs. Training powerful and robust AI models, especially deep neural networks, requires large amounts of high-quality and diverse training data that individual companies often do not have themselves, but are dispersed among different companies. Data sharing among multiple companies can help overcome this bottleneck but poses the risk of revealing sensitive business information or, when it comes to personal data, violating legal requirements such as the General Data Protection Regulation (GDPR). Federated learning, originally introduced for decentralized training of AI models on mobile devices, opens new opportunities for cross-enterprise collaboration by jointly training AI models without centralizing the underlying raw data.
Objective
The aim of the thesis is to understand the value creation potential of training AI models in federated networks for individual companies. To this end, factors influencing the individual value of collaboration should be identified and quantified.
Scope
- Literature review of company- and network-related factors influencing the individual value of collaboration in federated networks.
- Quantification of the identified factors in terms of the individual value of collaboration:
- Selection of one or more use cases, along with appropriate benchmark dataset(s) and algorithm(s).
- Definition of several federated network environments and different business conditions according to the identified influencing factors.
- Manipulation of federated network environments and business conditions to compare the performance of local/isolated models with the federated model and quantify the differences in the individual value of collaboration.
The thesis is supervised in cooperation with the Institute for Factory Automation and Production Systems (FAPS) (content person: Nils Thielen)
Language: German or English
Supervisor: Kristina Müller
Impact of inaccurate data in federated learning for artificial intelligence applications in manufacturing
Motivation
An important research focus for the surface mounted technology of electronics production is the use of artificial intelligence applications. This is due on the one hand to the high degree of automation and on the other hand to the large amount of available data. However, the data is often unbalanced due to the high yield, and certain defect patterns occur only in exceptional cases. This motivates the use of collaborative data utilization. Since companies do not want to share their data directly with potential competitors, federated learning lends itself. Here, however, it is critical that data quality is sufficient. Misclassified defects and different quality limits can thus lead to problems.
Objective
The goal of this thesis is to investigate a divergent data quality in federated learning on model quality. For this purpose, freely available benchmark data sets shall be chosen and analogies to production shall be defined. Based on this, both conventional and federated machine learning models have to be trained and tested. A crucial aspect is the targeted manipulation of the datasets to change the data quality.
Scope
This work may include the following key points:
- Preparation and aggregation of appropriate data sets with respect to federated learning
- Literature research
- Selection of suitable available data sets for the objective of this thesis
- Manipulation of the data set to change the quality of the data
- Development of machine learning models for benchmarking and federated learning
- Statistical and correlation analysis of the data
- Selection of suitable algorithms for model development
- Model training, testing, and optimization
- Validation and comparison of the results
- Analysis of the influence of deviating data quality
- Deriving further research potentials
- Consideration of the change from the model and data supplier perspective
The thesis is supervised in cooperation with the Institute for Factory Automation and Production Systems (FAPS) (content person: Nils Thielen)
Language: German or English
Supervisor: Kristina Müller
Mining market moves using natural language processing: From text to knowledge graph
Motivation
Knowledge graphs open up opportunities to represent complex relationships between entities for describing abstractly semantics of a real-world use case. To transform articles about competitor and market moves in the MedTech industry into a knowledge graph, companies and actions have to be detected. Natural language processing approaches can automize this task.
Tasks
- Short literature review to find related work and state of the art approaches
- Data collection: example code for fetching data is already available in C#, you can adapt and expand it
- Choose named entity recognition (NER) and relation extraction methods and apply them on your dataset
- Compare the results and recommend one method for the use case
Goal
Find the best performing NER and relation extraction approach for the dataset
Introductory Literature
Yang, S.M., Yoo, S.Y. and Jeong, O.R. (2020): “DeNERT-KG: Named Entity and Relation Extraction Model Using DQN, Knowledge Graph, and BERT”, in Applied Sciences (10). Doi: 10.3390/app10186429
Language: German or English
Supervisor: Annika Lurz
Partner: Siemens Healthineers
Formal Requirements
Before applying for a topic at our research group, please check the exam regulations of your specific study program regarding master thesis formal requirements (e.g., ECTS pre-requisite). If you have doubt please contact the responsible person of your study program directly.
For Master IIS students, there is no pre-requisite to do the master thesis with us. Nevertheless, we encourage students who may be interested to write their thesis with us, to first attend either our Business Intelligence lecture or one of our seminars.
Please note that students have to be matriculated while working on their master thesis.
Procedure
-
- Exposé
As soon as the topic is defined, you start working on the exposé. It consists of:- Motivation
- Research question
- Research design
- Expected results
- Preliminary list of references
- Registration
- Exposé
Once the exposé is approved, the topic can be officially registered. There are 6 months of official working time as per the rules of our faculty. This can be extended only after special circumstances usually requiring the submission of additional documents, such as doctor’s notice.
- Working Period
- You are responsible for organizing regular meetings with your supervisor.
- You should give a midterm presentation (10 to 15 minutes presentation plus 5 minutes questions) about halfway through your master thesis project.
- You should give a final presentation (15 minutes presentation plus 10 minutes questions).
- Submission
- You need to hand in two hard copies together with a CD or USB with all additional materials that have been used for the thesis (e.g., code, interview transcripts, references) directly to the examination office. Specific requirements for submission are usually stated with the registration confirmation from the examination office.
- The thesis should be approximately 25,000 words (60-80 pages).
Templates
- Word (english)
- We are currently working on providing a LaTeX template.
Ongoing Theses
- Trustworthy data sharing in healthcare: A blockchain-based concept and prototype (in cooperation with Swinburne University of Technology, Melbourne)
Supervisor: Pavlina Kröckel - The business and patients value of trusted data sharing within and between stakeholders in healthcare – lessons learned from explorative studies on private blockchains in healthcare (in cooperation with Swinburne University of Technology, Melbourne)
Supervisor: Pavlina Kröckel - The potential of NFTs in healthcare
Supervisor: Pavlina Kröckel, co-supervised with Prof. Mathias Kraus - AutoML platform at Siemens: Benchmark and market placement
Supervisor: Pavlina Kröckel - Open Government Data: Architecture and System Design for Enhancing Competitive Intelligence in the Medtech Industry
Supervisor: Annika Lurz - Medical Diagnostics Platform: Development of an App for Deep Learning Based Diagnostics of Blood Samples Measured with Deformability Cytometry
Supervisor: Annika Lurz
Partner: Max Planck Institut for the Science of Light - Entwicklung und Evaluierung eines Vorgehensmodells für die Digitale Transformation: Eine Fallstudie aus der Intralogistik in der Elektronikfertigung
Supervisor: Annika Lurz
Partner: Siemens Healthineers - Patent Analysis in the MedTech Industry Using Natural Language Processing Approaches
Supervisor: Annika Lurz - Agent-based Modeling of Dynamic Events: FAU Emergency Evacuation
Supervisor: Annika Lurz