Abstract
In this thesis, we study scalable and general purpose methods for mining frequent sequences that satisfy a given subsequence constraint. Frequent sequence mining is a fundamental task in data mining and has many real-life applications like information extraction, market-basket analysis, web usage mining, or session analysis. Depending on the underlying application, we are generally interested in discovering certain frequent sequences, which are described using subsequence constraints. There exists many tools and algorithms for this task, however, they are not sufficiently scalable to deal with large amounts of data that may arise in applications and are generally not extensible across range of applications. We propose scalable, distributed sequence mining algorithms that target MapReduce. Our work builds on MG-FSM, which is a distributed framework for frequent sequence mining. We propose novel algorithms that improve and extend the basic MG-FSM framework to efficiently support traditional subsequence constraints that arise in applications. Additionally, we show that many subsequence constraints—including and beyond the traditional ones considered in literature—can be unified in a single framework. A unified treatment allows researchers to study jointly many types of subsequence constraints (instead of each one individually) and helps to improve usability of pattern mining systems for practitioners. To this end, we propose a general purpose framework that provides a set of simple and intuitive ``pattern expressions’’, which allows to describe any subsequence constraint of interest and explore algorithms for efficiently mining frequent subsequences under such general constraints. Our experimental study on real-world datasets indicates that our proposed algorithms are scalable and effective across wide range of applications.
