Characteristics and commonalities: differentially describing datasets with insightful patterns

Sebastian Dalleiger

Abstract: Empirical science revolves around gaining insights from complex data. With the advent of computational science, increasingly more, larger, and richer datasets are becoming available to expand our scientific knowledge. However, the analysis of these datasets by domain experts is often impaired by a lack of suitable computational tools. In particular, there is a shortage of methods identifying insightful patterns, i.e., sets of strongly associated feature values that are informative, contrasting, probabilistically sound, statistically sound, and discoverable using scalable algorithms. This thesis leverages ideas and concepts from pattern-set mining, maximum-entropy modeling, statistical testing, and matrix factorization to develop methods for discovering insightful patterns.