"Statistical graph clustering"
Event details
| Date | 26.01.2017 |
| Hour | 10:00 › 11:00 |
| Speaker | Prof. Emmanuel Abbe (Princeton University) |
| Location | |
| Category | Conferences - Seminars |
Clustering is at the core of unsupervised machine learning and network data analysis. Despite significant advances at the methodological level, establishing fundamental limits remains a challenge. Our approach is based on a statistical framework to graph clustering (community detection), starting with stochastic block models. In this context, we obtain sharp characterizations of when clusters estimations are possible or not. These are obtained from two proof techniques: local to global hypothesis testing and inference on branching processes. We further establish the physicists’ conjecture on the emergence of a gap between the statistical and computational thresholds, using a sampling method. This underlines the importance of bridging statistical and computational methods in this field, likely to be a recurrent theme in data sciences. Data implementations, spectral methods and connections to Ramanujan graphs will also be discussed.
Practical information
- Informed public
- Free
Organizer
- Prof. Philippe Michel
Contact
- marcia gouffon