Inferring graph structure from signal observations
Event details
| Date | 25.08.2016 |
| Hour | 11:00 › 13:00 |
| Speaker | Hermina Petric Maretic |
| Location | |
| Category | Conferences - Seminars |
EDIC Candidacy Exam
Exam President: Prof. Pierre Vandergheynst
Thesis Director: Prof. Pascal Frossard
Co-examiner: Prof. Daniel Kressner
Background papers:
Learning Laplacian Matrix in Smooth Graph Signal Representations, by X. Dong, et al.
Fitting a Graph to Vector Data, by S.Daitch, et al.
Efficient Dimensionality Reduction for High-Dimensional Network Estimation, by S. Celik, et al.
Abstract
Network-structured data appears naturally in a large and constantly increasing number of domains, which makes its analysis crucial. However, in order to successfully process data on graphs, it is essential to place them on a meaningful graph structure. As these structures are often not known or uniquely defined, we encounter the problem of inferring relevant graphs from signals.
In this proposal, we discuss three different approaches to the problem of graph learning. We first examine a more traditional approach from the machine learning community, modelling network structure to get further insight in solving machine learning problems. We then take a look at a signal processing approach, based on an extension of the classical factor analysis model, leveraging the connection of the graph Fourier transform to its topology. Finally, we consider an approach relying on a probabilistic model to learn large and highly structured graphs. We finish with a short overview of our current work and a discussion of potential further directions.
Exam President: Prof. Pierre Vandergheynst
Thesis Director: Prof. Pascal Frossard
Co-examiner: Prof. Daniel Kressner
Background papers:
Learning Laplacian Matrix in Smooth Graph Signal Representations, by X. Dong, et al.
Fitting a Graph to Vector Data, by S.Daitch, et al.
Efficient Dimensionality Reduction for High-Dimensional Network Estimation, by S. Celik, et al.
Abstract
Network-structured data appears naturally in a large and constantly increasing number of domains, which makes its analysis crucial. However, in order to successfully process data on graphs, it is essential to place them on a meaningful graph structure. As these structures are often not known or uniquely defined, we encounter the problem of inferring relevant graphs from signals.
In this proposal, we discuss three different approaches to the problem of graph learning. We first examine a more traditional approach from the machine learning community, modelling network structure to get further insight in solving machine learning problems. We then take a look at a signal processing approach, based on an extension of the classical factor analysis model, leveraging the connection of the graph Fourier transform to its topology. Finally, we consider an approach relying on a probabilistic model to learn large and highly structured graphs. We finish with a short overview of our current work and a discussion of potential further directions.
Practical information
- General public
- Free
Contact
- Cecilia Chapuis EDIC