Learning Bayesian Networks Through Knowledge Reduction
Author
Hernandez, Jorge Pablo Cordero
Term
10. Term
Education
Publication year
2007
Abstract
At lære strukturen i bayesianske netværk ud fra data bliver hurtigt beregningsmæssigt uoverkommeligt, når antallet af variabler vokser. En udbredt løsning er en dele-og-hersk-strategi, hvor et stort problem deles i mindre, håndterbare dele. Denne afhandling giver en generel måde at gøre dette på ved at introducere en klasse af modeller kaldet Bayesian network knots. Disse modeller opdeler eksplicit variablerne i lokale komponenter i netværket. Med udgangspunkt i denne idé foreslår vi Overlapping Expansion Learning (OSL)-algoritmen til at lære bayesianske netværk og undersøger, hvordan attributklyngedannelse, altså gruppering af variabler med lignende adfærd, påvirker læringen. Eksperimenter viser, at OSL er meget konkurrencedygtig. Vi introducerer også en ny algoritme til attributklyngedannelse, Star Discovery (SD), som finder grupper af variabler med højere ydeevne end flere andre tilgange. Samlet set hjælper disse bidrag med at gøre læring af store bayesianske netværk mere gennemførlig.
Learning the structure of Bayesian networks from data quickly becomes computationally intractable as the number of variables grows. A common remedy is a divide-and-conquer strategy that splits a large problem into smaller, manageable parts. This thesis provides a general way to apply this strategy by introducing a class of models called Bayesian network knots. These models explicitly partition the variables into local components within the network. Building on this idea, we propose the Overlapping Expansion Learning (OSL) algorithm for learning Bayesian networks and examine how attribute clustering, that is, grouping variables with similar behavior, affects learning. Experiments show that OSL is highly competitive. We also introduce a new attribute clustering algorithm, Star Discovery (SD), which discovers groups of variables with higher performance than several other approaches. Together, these contributions help make learning large Bayesian networks more feasible.
[This abstract was generated with the help of AI]
Documents