Here you can find the currently available thesis and project topics offered by our research group.
If you would like more information about a listed topic, please contact the corresponding supervisor(s).

Open Theses

Large Language Models and Knowledge Graphs

Type: Bachelor’s or Master’s Thesis

Objective: You will work on elaborating the use of knowledge graphs for large language models (LLMs). In particular, we are interested in summarizing current work focusing on this topic, discussing the limitations of current work, and suggesting improvements. Alternatively, your job may utilize LLMs together with knowledge graphs for coming up with a solution to a practical problem, such as diagnosis or configuration.

Core content of the work:

• Elaboration on the current combination of LLMs and Knowledge Graphs
  • Providing a survey or a lightweight structured review of existing literature
• Evaluation of the current approaches
  • What are existing applications?
  • What can be done and what not?
  • What are the limitations?

Technical details:

• Language: Python
• Tools/frameworks: scikit-learn, knowledge-graph frameworks

Contact: wotawa@tugraz.at

ASP for Testing

Type: Bachelor’s or Master’s Thesis

Objective: There is a paper dealing with utilizing Answer Set Programming (ASP) for generating combinatorial test suites. Unfortunately, the implementation mentioned in this paper is not available. Hence, your job is to implement a tool that utilizes ASP for generating combinatorial test suites.

Core content of the thesis:

• Implementing a combinatorial test suite generation algorithm in ASP
• Evaluating the test suite generation approach, including a comparison with existing tools
  • Test case generation time
  • Quality of generated test suites (e.g., smaller test suites are preferred)

Technical details:

• Language: Python
• Tools/frameworks: clingo (ASP), ACTS

Contact: wotawa@tugraz.at

Comparison of Explainability Methods for ML-based Fault Diagnosis in Building Automation Systems

Type: Master’s Thesis

Objective: You will apply various methods for explaining machine learning models to tried-and-tested classification models (e.g., decision tree, random forest, multi-layer perceptron). The focus is on comparing explainability in terms of comprehensibility, visualizability, computational effort, and model robustness. There is a public dataset on heating systems that we use for this purpose (Nature Dataset).

Core content of the work:

• Application and comparison of explainability methods
  • LIME (Local Interpretable Model-agnostic Explanations)
  • SHAP (SHapley Additive exPlanations)
  • Feature Importance (model-dependent feature weighting)
  • Rule Extraction (for decision trees)
• Evaluation according to criteria such as
  • Interpretability / comprehensibility
  • Visualization options
  • Runtime behavior
  • Sensitivity to model variants

Technical details:

• Language: Python
• Tools/frameworks: scikit-learn, shap, lime, matplotlib

Contact: wotawa@tugraz.at, rkoitz-hristov@tugraz.at

Hyperparameter Tuning for ML Models for Fault Diagnosis in Building Automation Systems

Type: Master’s Thesis

Objective: The aim of this thesis is to systematically investigate how targeted hyperparameter tuning affects the accuracy, robustness, and behavior of machine learning models for fault diagnosis in building automation systems. Different optimization strategies will be implemented and compared based on performance and efficiency. There is a public dataset on heating systems that we use for this purpose (Nature Dataset). 

Core content of the thesis:

• Tuning strategies and model comparison
  • Grid search
  • Random search
  • Bayesian optimization
  • Combinatorial designs (aka “Combinatorial Testing”)
• Evaluation according to criteria such as
  • Training/test time
  • Metrics (e.g., F1 score, precision, recall)

Technical details:

• Language: Python
• Tools/frameworks: scikit-learn, Optuna

Contact: wotawa@tugraz.at, rkoitz-hristov@tugraz.at

Ablation Study: Influence of Reduced Sensor Technology on the Diagnostic Performance of ML Models in HVAC Systems

Type: Bachelor’s or Master’s Thesis

Objective: The goal of this thesis is to study how reducing the number of available sensor signals (feature subsets) influences the diagnostic performance of machine learning models in HVAC systems. There is a public dataset on heating systems that we use for this purpose (Nature Dataset). 

Core content of the work:

• Feature subset
  • Comparison of different feature subsets and their effect on diagnostic accuracy/robustness
  • Random subsets (to establish variance)
  • Domain-driven subsets (e.g., temperature only, flow only)
  • Importance-driven subsets (SHAP values)
  • Use of explainability methods to evaluate feature importance
• Evaluation according to criteria such as
  • Training/testing time
  • Metrics (e.g., F1 score, precision, recall)
  • Robustness
  • Feature importance (SHAP)

Technical details:

• Language: Python
• Tools/frameworks: scikit-learn, shap, matplotlib

Contact: wotawa@tugraz.at, rkoitz-hristov@tugraz.at