Learning Bayesian Networks Using Fast Heuristics

This thesis addresses score-based learning of Bayesian networks from data using a few fast heuristics. The algorithmic implementation of the heuristics is able to learn size 30-40 networks in seconds and size 1000-2000 networks in hours. Two algorithms, which are devised by Scanagatta et al. and dubbed Independence Selection and Acyclic Selection OBS have the capacity of learning very large Bayesian networks without the liabilities of the traditional heuristics that require maximum in-degree or ordering constraints. The two algorithms are respectively called Insightful Searching and Acyclic Selection Obeying Boolean-matrix Sanctioning (acronym ASOBS) in this thesis. This thesis also serves as an expansion of the work of Scanagatta et al. by revealing a computationally simple ordering strategy called Randomised Pairing Greedy Weight (acronym RPGw) that works well as an adjunct along with ASOBS with corresponding experiment results, which show that ASOBS was able to score higher and faster with the help of RPGw. Insightful Searching, ASOBS, and RPGw together form a system that learns Bayesian networks from data very fast