Characterising the Di↵erence and the Norm between Sequence Databases

In pattern set mining we are after a small set of patterns that together are characteristic for the data at hand. In this paper we consider the problem of characterizing not one, but a set of sequence databases, such as a collection of articles or the chapters of a book. Our main objective is to find a set of patterns that captures the individual features of each database, while also finding shared characteristics of any subset of the data we are interested in. We formulate this problem in terms of MDL, and propose SqsNorm, an e cient algorithm to extract high quality models directly from data. We devise a heuristic to quickly find and evaluate candidates, as well as a model that takes shared and singular patterns into account. Experiments on synthetic data confirm that SqsNorm ably reconstructs the ground truth model. Experiments on text data, including a set of speeches, several book chapters, and a collection of songs, show that SqsNorm discovers informative, non-redundant and easily interpretable pattern sets that give clear insight in the data.