3rd International Workshop on Similarity-Based Pattern Analysis and Recognition

Invited speakers


Finding Endogenously Formed Communities

Speaker: Nina Balcan, Carnegie Mellon University, USA.

Nina Balcan

An important unsupervised learning task which has received significant recent interest is identifying overlapping clusters, or communities, in networks ranging from professional contact networks to citation networks to product purchasing networks. While many heuristics and optimization criteria have been proposed, a lot of the previous work has disallowed natural communities such as those containing highly popular nodes, or have not given general guarantees on the computation time needed to find all overlapping communities meeting certain criteria. In this work, we develop effective methods for identifying natural self-determined communities in social networks and in more general affinity systems. These communities have the property that their members collectively prefer each other to anyone else outside the community. By contrast to previous work, our new formalization leads to discovering natural types of communities and enabled us to design efficient algorithms for identifying all such communities. Furthermore, for interesting settings of the parameters, we also provide a local algorithm with a strong stochastic performance guarantees that can find a community in time nearly linear in the of size the community (as opposed to the size of the network).

Bio sketch: Maria-Florina Balcan is an Associate Professor in the School of Computer Science at Carnegie Mellon University. Her main research interests are machine learning, computational aspects in economics and game theory, and algorithms. Her honors include the CMU SCS Distinguished Dissertation Award, an NSF CAREER Award, a Microsoft Faculty Research Fellowship, a Sloan Research Fellowship, and several paper awards. She was a Program Committee Co-chair for COLT 2014 , and is currently a board member of the International Machine Learning Society and a Program Committee Co-chair for ICML 2016.


Structure metric learning learning for prototype-based models

Speaker: Barbara Hammer, Bielefeld University, Germany.

hammer small

Prototype-based learning techniques such as learning vector quantization (LVQ) enjoy a wide popularity due to their intuitive training and model interpretability. Applications include biomedical data analysis, image classification, or fault detection in technical systems. We will focus on modern variants of LVQ which are based on cost functions and which can be complemented with a clear learning theoretical foundation. Within the talk, we will focus on two recent extensions of these techniques which are of interest as soon as data become more complex: Classical LVQ has been proposed for vectorial data only. So the question is how to extend this technology if data are given in terms of proximities only? It turns out that the cost-function-based point of view of LVQ opens the way for a very generic extension of LVQ towards similarities, dissimilarities and kernels. In the talk, we will present relational and kernel extensions of LVQ and we will show in how far these different variants are specific instances of one general framework. LVQ crucially depends on the metric which is used to compare data; usually, this metric comes from some parametrized form, and the exact choice of the metric parameters can be crucial for the success. This principle is well established for vectorial LVQ, and often referred to as relevance or matrix learning. We show to extend this principle towards structure metric learning for proximity variants of LVQ; there, we focus on alignment distances and an autonomous adaptation of the underlying scoring function.

Slides of the talk: pdf

Bio sketch: Barbara Hammer received her Ph.D. in Computer Science in 1995 and her venia legendi in Computer Science in 2003, both from the University of Osnabrueck, Germany. From 2000-2004, she was chair of the junior research group 'Learning with Neural Methods on Structured Data' at University of Osnabrueck before accepting an offer as professor for Theoretical Computer Science at Clausthal University of Technology, Germany, in 2004. Since 2010, she is holding a professorship for Theoretical Computer Science for Cognitive Systems at the CITEC cluster of excellence at Bielefeld University, Germany. Several research stays have taken her to Italy, U.K., India, France, the Netherlands, and the U.S.A. Her areas of expertise include hybrid systems, self-organizing maps, clustering, and recurrent networks as well as applications in bioinformatics, industrial process monitoring, or cognitive science. She has been chairing the IEEE CIS Technical Committee on Data Mining in 2013 and 2014, and she is chair of the Fachgruppe Neural Networks of the GI and vice-chair of the GNNS. She has published more than 200 contributions to international conferences / journals, and she is coauthor/editor of four books.


Non-parametric Bayesian Modeling of Relational Data

Speaker: Morten Mørup, Technical University of Denmark, Denmark.

Relational data/complex networks characterizing relationships between entities emerge in practically all fields of research. A common aim modeling these data is to extract structure at the level of groups from measures of similarities/links between entities. This talk will focus on non-parametric Bayesian models for relational data. Notably, these models devise statistical processes for generating data, are able to determine model complexity as part of the inference, account for parameter uncertainty, and can be used to predict similarities/links. A starting point will be the infinite relational model that is a non-parametric extension of the stochastic block model. Extensions will be treated including the modelling of degree heterogeneity, community structure and hierarchical structure. Applications to the modeling of functional and structural brain connectivity as well as social networks will be demonstrated.

Slides of the talk: pdf

Bio sketch: Morten Mørup is associate professor at the Section for Cognitive Systems at DTU Compute, Technical University of Denmark. His research focuses on machine learning, Bayesian modeling and neuroimaging where he primarily research unsupervised learning strategies for modeling neuroimaging data, networks and multi-way data sets. He is associate editor of IEEE Transactions on Signal Processing, and member of the Machine Learning for Signal Processing Technical Committee of the IEEE Signal Processing Society. Morten Mørup has received several awards including an award from His Royal Highness the Prince's Foundation in 2013, a Lundbeck Foundation Fellowship in 2012, a best teacher award at DTU Informatics 2011, a thesis award from Direktør Peter Gorm-Petersens Mindelegat in 2008, and an Elite Research travel scholarship from the Danish Ministry of Science in 2007.