Claudia Plant
Data Mining Research Group
University of Vienna

Christian Boehm
Data Mining
Ludwig Maximilian University of Munich

"Mining Patterns that Compress"

Tuesday, September 18, 2018
3:30-4:30 p.m.
499 Dirac Science Library

Abstract:

Exploratory data analysis aims at detecting any kind of interesting structure or pattern in data. How to measure the success of data-driven exploration? This is a challenging question because we want to explore data without restricting ourselves to specific types of patterns and we do not assume that a labeled training data set is available. Data compression gives an intuitive and powerful answer. If data contains any kind of structure which we can find by an algorithm and represent by some kind of probabilistic model, we can exploit our knowledge about this structure for effectively compressing the data. The first part of this talk will introduce this basic idea taking exploratory graph mining as an example task. We will then survey some of our recent work applying this idea to various types of tasks, e.g., clustering and dimensionality reduction on different types of data. In the second part of this talk we will focus on highly efficient data mining methods for exploratory analysis of massive data. An outlook on ongoing work and open questions including non-parametric modelling and the large parameter spaces of deep nets will conclude the talk.