Education

Theses

We keep a running list of possible topics for bachelor and master theses in the areas of pattern recognition. In addition, we may also offer topics for bachelor theses in the area of human-computer interaction. If you are interested, please contact us.

TitleLevelTypeContact
New Data Sets for Graph-Based Machine LearningMA/BA in the area of PRHaving data sets from diverse applications is a crucial prerequisite for doing research in the field of pattern recognition and machine learning. The goal of this project is to compile and preprocess at least four data sets that are readily applicable in the algorithmic frameworks used in our research group. That is, we are particularly interested in graph-based data. An experimental evaluation using standard algorithms stemming from the arsenal of graph-based machine learning methods is also an important pillar of the work.kaspar.riesen@inf.unibe.ch
Graph reduction with graph coarseningMA/BA in the area of PRDiverse graph reduction methods have already been proposed in the literature and it is an essential task to speed up graph matching algorithms.
This project would try to use new Graph Coarsening (GC) methods to reduce the graphs.
GC consists of iteratively running a graph community detection and merging the nodes in each individual community.
The coarsening method is run until the good number of nodes remain in the graph.
A thorough comparison with node sampling methods (using centrality measures) will be done to assess the efficiency of the reduction method.
kaspar.riesen@inf.unibe.ch / anthony.gillioz@inf.unibe.ch
Improving the conservation of information during graph reduction processesMA/BA in the area of PRWorking with reduced graphs has multiple advantages, in particular, it can speed up graph matching methods.
Unfortunately, reducing graphs is not a lossless information process.
When using a node sampling graph reduction method, the nodes are selected based on their importance in the graph structure.
The nodes that are not kept in the reduced version of the graph are deleted and their internal information is lost.
The idea of this project would be to encode the node information and dilute the information of the deleted nodes into the remaining nodes.
kaspar.riesen@inf.unibe.ch /
anthony.gillioz@inf.unibe.ch
How important node information is when performing graph matching?MA/BA in the area of PRGraph matching algorithms are using node information to improve matching performance.
Unfortunately, it is not clear how important that information is in the graph matching process.
In this project, we try to find out how important the node information is compared to the graph structure.
So, we propose to run two different sets of experiments (w/ and w/o node label) on various graph matching algorithms and see if any difference appears.
anthony.gillioz@inf.unibe.ch
Similarity of Graphs: Direct Product matching-graphMA/BA in the area of PR The basic idea is to formalize the stable cores of individual classes of graphs, discovered during intra-class matching, by means of so-called matching-graphs.
A direct product graph is a graph that contains each possible node combination of two given graphs as well as corresponding edges. This graph is for example used during the random walk kernel.
In this project, we aim at creating an adapted version of this direct product graph and try to use it to improve distance-based graph classification accuracy.
kaspar.riesen@inf.unibe.ch / mathias.fuchs@inf.unibe.ch
Graph Augmentation using Matching-Graphs for graph neural networksMA/BA in the area of PR Matching-graphs are small graphs that represent the similarity between two graphs based on a given matching between them.
Based on a given training set of graphs, a lot of matching-graphs can be created. It is known that especially classifiers like neural networks are extremely data-dependent.
In this project, we aim at generating a lot of matching-graphs with the idea of augmenting a given training set. This augmented training set should then be tested with several graph neural network-based classifiers.
kaspar.riesen@inf.unibe.ch / mathias.fuchs@inf.unibe.ch
Gulf of Evaluation/Execution — eine Sammlung und Analyse von Use CasesBA in HCI (german)Das 7-Stufen Modell von Norman gilt als einflussreichstes Modell in der HCI. In dieser Arbeit sollen systematisch Beispiele aus Echt-Welt-Anwendungen gesammelt und mit dem Modell von Norman analysiert werden. Insbesondere soll mit ca. 7 verschiedenen Beispielen gezeigt werden, wie ein grosser Gulf of Execution bzw. Evaluation die Usability eines digitalen Produktes drastisch verschlechtert wird. Zudem erarbeiten Sie in Ihrer Arbeit einen ersten Entwurf eines Fragebogens, der Software Entwicklern helfen kann, grosse Gulfes semi-automatisch zu identifizieren.kaspar.riesen@inf.unibe.ch

Courses

TitleDescriptionSemester
Mensch-Maschine SchnittstelleHS 21
Programmieren für NaturwissenschaftenHS 21
Programmierung 1HS 21