Biclustering is an increasingly popular technique used for unsupervised data analysis in several domains, from gene expression profiling to user preferences and recommendation. Despite this, most algorithms still depend mainly on exhaustive approaches to obtain quality results. Since Reinforcement Learning has been used to tackle combinatorial optimization problems with great success, we explore its application to improve on current biclustering (or general n-clustering) heuristics and provide a way of solving this task more efficiently.
This thesis relates to more fundamental research, focusing on a methodological approach to solving a general problem that can be applied in many different domains. The main results include a new set of tools to deal with three-way data, ranging from healthcare applications (patients-features-time) to social or usage data (users-preferences-time or users-users-time).
Author: Pedro Cotovio
Type: MSc thesis
Partner: FCUL – Faculdade de Ciências da Universidade de Lisboa