Variational inference for fully bayesian hierarchical linear models

dc.contributor.advisorSosa Martinez, Juan Camilo
dc.contributor.authorParra Aldana, Cristian Camilo
dc.contributor.cvlacSosa Martinez, Juan Camilo [0001359814]
dc.contributor.googlescholarSosa Martinez, Juan Camilo [armR6koAAAAJ]
dc.contributor.orcidSosa Martinez, Juan Camilo [0000000174324014]
dc.contributor.researchgateSosa Martinez, Juan Camilo [Juan-Sosa-10]
dc.contributor.scopusSosa Martinez, Juan Camilo [80099012]
dc.date.accessioned2026-02-11T15:21:22Z
dc.date.available2026-02-11T15:21:22Z
dc.date.issued2025
dc.descriptionilustraciones a colro, diagramasspa
dc.description.abstractLos modelos lineales jerárquicos bayesianos ofrecen un marco natural para capturar estructuras anidadas y de agrupamiento en los datos. La estimación clásica mediante Cadenas de Markov Monte Carlo (MCMC) proporciona distribuciones posteriores bien calibradas, pero con un alto costo computacional, a menudo prohibitivo en escenarios de gran escala o alta dimensión. En contraste, la Inferencia Variacional (VI) y su variante estocástica (SVI) han surgido como alternativas eficientes, basadas en la optimización y no en el muestreo. Su rapidez, sin embargo, se logra a expensas de la calidad de la aproximación, especialmente en contextos jerárquicos cuando la separación intrinseca de los grupos no está marcada a priori. Esta tesis evalúa críticamente las ventajas y desventajas de ambos paradigmas a lo largo de diferentes niveles de complejidad: el Modelo de Regresión Lineal (LRM), el Modelo Jerárquico (HLRM) y el Modelo Jerárquico con Agrupamiento (CHLRM). Los estudios de simulación y la aplicación a datos reales muestran que VI y SVI reproducen efectos globales y patrones de agrupamiento con un tiempo de cómputo mínimo, pero distorsionan sistemáticamente las dependencias posteriores y generan valores inestables en criterios de información como WAIC y DIC. El aporte de este trabajo consiste en clarificar la diferencia de alcance entre la inferencia basada en muestreo y la basada en optimización, destacando los contextos en que los métodos variacionales pueden actuar como sustitutos prácticos de MCMC y aquellos en los que sus limitaciones son críticas. Más allá de la comparación metodológica, la tesis cumple también un papel pedagógico al hacer accesible la computación bayesiana avanzada, y señala futuras extensiones bajo el mismo marco variacional hacia modelos lineales generalizados (GLM) y otros miembros de la familia exponencial. (Texto tomado de la fuente)spa
dc.description.abstractBayesian hierarchical linear models provide a natural framework to capture nested and clustered structures in data. Classical estimation via Markov Chain Monte Carlo (MCMC) delivers well-calibrated posterior distributions but is computationally demanding, often prohibitive in high-dimensional or large-scale settings. In contrast, Variational Inference (VI) and Stochastic Variational Inference (SVI) have emerged as efficient alternatives, relying on optimization rather than sampling. Their tractability, however, comes at the cost of approximation quality, especially in hierarchical contexts where the intrinsic separation of groups is not marked a priori. This thesis critically evaluates the advantages and disadvantages of both paradigms across increasing levels of complexity: the Linear Regression Model (LRM), the Hierarchical Linear Regression Model (HLRM), and the Clustered HLRM (CHLRM). Simulation studies and an application to real data show that VI and SVI reproduce global regression effects and clustering patterns with minimal runtime, but systematically distort posterior dependencies and yield unstable information criteria such as WAIC and DIC. The contribution of this work lies in clarifying the difference in scope between sampling-based and optimization-based inference, highlighting contexts where variational methods can act as practical surrogates to MCMC, and where their limitations are critical. Beyond the methodological comparison, the thesis also serves a pedagogical role, making advanced Bayesian computation accessible, and points to future extensions under the same variational framework to generalized linear models (GLMs) and other members of the exponential family.eng
dc.description.degreelevelMaestría
dc.description.degreenameMagíster en Ciencias - Estadística
dc.description.methodsBuilding on the theoretical foundations of Bayesian inference and variational methods, this chapter develops the methodological framework of the thesis. We formalize the Linear Regression Model (LRM), the Hierarchical Linear Regression Model (HLRM), and the Clustering HLRM (CHLRM), specifying their priors and posterior structures. These models provide the ground on which we implement and compare MCMC, Variational Inference (VI), and Stochastic Variational Inference (SVI), thus establishing the link between theory and application that guides the subsequent empirical analysis.
dc.description.notes
dc.description.researchareaEstadística Bayesiana
dc.description.technicalinfoAll code developed for this thesis is publicly available in a GitHub repository at https://github.com/ccparra/VariationalBayes-HLRM/. The implementation follows an object-oriented design using the R6 library in R, providing the reader with reusable classes and methods to freely apply and extend the algorithms presented in this work.
dc.format.extentxv, 102 páginas
dc.format.mimetypeapplication/pdf
dc.identifier.instnameUniversidad Nacional de Colombiaspa
dc.identifier.reponameRepositorio Institucional Universidad Nacional de Colombiaspa
dc.identifier.repourlhttps://repositorio.unal.edu.co/spa
dc.identifier.urihttps://repositorio.unal.edu.co/handle/unal/89494
dc.language.isoeng
dc.publisherUniversidad Nacional de Colombia
dc.publisher.branchUniversidad Nacional de Colombia - Sede Bogotá
dc.publisher.facultyFacultad de Ciencias
dc.publisher.placeBogotá, Colombia
dc.publisher.programBogotá - Ciencias - Maestría en Ciencias - Estadística
dc.relation.referencesAmari, S.-i. (1998). Natural gradient works efficiently in learning. Neural Computation, 10(2):251– 276
dc.relation.referencesBeal, M. J. (2003). Variational algorithms for approximate Bayesian inference. University of London, University College London (United Kingdom)
dc.relation.referencesBishop, C. M. (2007). Pattern Recognition and Machine Learning (Information Science and Statistics). Springer, 1 edition
dc.relation.referencesBlei, D. M. and Jordan, M. I. (2006). Variational inference for dirichlet process mixtures. Bayesian Anal., 1(1):121–143
dc.relation.referencesBlei, D. M., Kucukelbir, A., and McAuliffe, J. D. (2016). Variational inference: A review for statisticians. Journal of the American Statistical Association, 112:859 – 877
dc.relation.referencesBrown, L. D. (1986). Fundamentals of statistical exponential families with applications in statistical decision theory. Lecture Notes-Monograph Series, 9:i–279
dc.relation.referencesChristensen, R., Johnson, W. O., Branscum, A. J., and Hanson, T. E. (2010). Bayesian ideas and data analysis. Chapman & Hall/CRC Texts in Statistical Science. CRC Press, Boca Raton, FL
dc.relation.referencesCrawley, M. J. (2012). The R Book. Wiley-Blackwell, Hoboken, NJ, 2 edition
dc.relation.referencesDiaconis, P. and Ylvisaker, D. (1979). Conjugate priors for exponential families. Ann. Stat., 7(2):269–281
dc.relation.referencesGamerman, D. and Lopes, H. F. (2006). Markov chain Monte Carlo. Chapman & Hall/CRC Texts in Statistical Science. Chapman & Hall/CRC, Philadelphia, PA, 2 edition
dc.relation.referencesGelman, A., Carlin, J., Stern, H., Dunson, D., Vehtari, A., and Rubin, D. (2014). Bayesian Data Analysis. CRC Press, Florida
dc.relation.referencesGelman, A. and Hill, J. (2007). Data Analysis Using Regression and Multilevel/Hierarchical Models. Cambridge, New York
dc.relation.referencesGhahramani, Z. and Beal, M. (2000). Propagation algorithms for variational bayesian learning. In Leen, T., Dietterich, T., and Tresp, V., editors, Advances in Neural Information Processing Systems, volume 13. MIT Press
dc.relation.referencesHoff, P. D. (2009). A First Course in Bayesian Statistical Methods. Springer, Seattle WA
dc.relation.referencesHoffman, M., Blei, D. M., Wang, C., and Paisley, J. (2013). Stochastic variational inference. Journal of Machine Learning Research, 14 No. 4:1303–1347
dc.relation.referencesKass, R. E. and Wasserman, L. (1995). A reference bayesian test for nested hypotheses and its relationship to the schwarz criterion. J. Am. Stat. Assoc., 90(431):928–934
dc.relation.referencesKingma, D. P. and Welling, M. (2022). Auto-encoding variational bayes
dc.relation.referencesKullback, S. and Leibler, R. A. (1951). Information and sufficiency. The Annals of Mathematical Statistics, 22:79–86
dc.relation.referencesMurphy, K. P. (2012). Machine Learning: A Probabilistic Perspective. The MIT Press
dc.relation.referencesOrmerod, J. T. and Wand, M. P. (2010). Explaining variational approximations. The American Statistician, 64(2):140–153
dc.relation.referencesParra, C. (2025). Variational Inference in Hierarchical Linear Regression Models. https://github. com/ccparra/VariationalBayes-HLRM/
dc.relation.referencesRanganath, R., Gerrish, S., and Blei, D. (2014). Black box variational inference. Technical report, Princeton University
dc.relation.referencesRegueiro, P., Rodríguez, A., and Sosa, J. (2024). Stochastic gradient variational bayes in the stochastic blockmodel
dc.relation.referencesRobbins, H. and Monro, S. (1951). A stochastic approximation method. Annals of Mathematical Statistics, 22:400–407
dc.relation.referencesSosa, J. and Aristizabal, J. (2022). Some developments in bayesian hierarchical linear regression modeling. Revista Colombiana de Estadística - Applied Statistics, 45:231–255
dc.relation.referencesWainwright, M. J. and Jordan, M. I. (2007). Graphical models, exponential families, and variational inference. Found. Trends® Mach. Learn., 1(1-2):1–305
dc.relation.referencesWang, H., Bhattacharya, A., Pati, D., and Yang, Y. (2022). Structured variational inference in bayesian state-space models. In Camps-Valls, G., Ruiz, F. J. R., and Valera, I., editors, Proceedings of The 25th International Conference on Artificial Intelligence and Statistics, volume 151 of Proceedings of Machine Learning Research, pages 8884–8905. PMLR
dc.relation.referencesWatanabe, S. (2012). A widely applicable bayesian information criterion. CoRR, abs/1208.6338
dc.relation.referencesZellner, A. (1986). On assessing prior distributions and Bayesian regression analysis using g-prior distributions, pages 233–243. NorthHolland, Amsterdam
dc.relation.referencesZetterström, V. (2021). Variational bayes as a computer intensive method for bayesian regression. Master’s thesis, Uppsala University
dc.relation.referencesZhang, C., Butepage, J., Kjellstrom, H., and Mandt, S. (2018). Advances in variational inference
dc.rights.accessrightsinfo:eu-repo/semantics/openAccess
dc.rights.licenseAtribución-NoComercial-CompartirIgual 4.0 Internacional
dc.rights.urihttp://creativecommons.org/licenses/by-nc-sa/4.0/
dc.subject.ddc000 - Ciencias de la computación, información y obras generales::004 - Procesamiento de datos Ciencia de los computadores
dc.subject.ddc000 - Ciencias de la computación, información y obras generales::005 - Programación, programas, datos de computación
dc.subject.ddc000 - Ciencias de la computación, información y obras generales::006 - Métodos especiales de computación
dc.subject.ddc510 - Matemáticas::519 - Probabilidades y matemáticas aplicadas
dc.subject.lembDECISIONES ESTADISTICASspa
dc.subject.lembStatistical decisioneng
dc.subject.lembTEORIA BAYESIANA DE DECISIONES ESTADISTICASspa
dc.subject.lembBayesian statistical decision theoryeng
dc.subject.lembMODELOS LINEALES (ESTADISTICA) PROCESAMIENTO DE DATOSspa
dc.subject.lembLinear models (statistics) - data processingeng
dc.subject.lembPROCESOS DE MARKOVspa
dc.subject.lembMarkov processeseng
dc.subject.lembMETODO DE MONTECARLOspa
dc.subject.lembMonte carlo methodeng
dc.subject.proposalBayesian hierarchical modelseng
dc.subject.proposalVariational Inferenceeng
dc.subject.proposalStochastic Variational Inferenceeng
dc.subject.proposalExponential familyeng
dc.subject.proposalEvidence Lower Boundeng
dc.subject.proposalBayesian computationeng
dc.subject.proposalModelos jerárquicos bayesianosspa
dc.subject.proposalInferencia Variacionalspa
dc.subject.proposalInferencia Variacional Estocásticaspa
dc.subject.proposalFamilia exponencialspa
dc.subject.proposalComputación bayesianaspa
dc.subject.proposalELBOspa
dc.titleVariational inference for fully bayesian hierarchical linear modelseng
dc.title.translatedInferencia variacional para modelos lineales jerárquicos completamente bayesianosspa
dc.typeTrabajo de grado - Maestría
dc.type.coarhttp://purl.org/coar/resource_type/c_bdcc
dc.type.coarversionhttp://purl.org/coar/version/c_ab4af688f83e57aa
dc.type.contentText
dc.type.driverinfo:eu-repo/semantics/masterThesis
dc.type.redcolhttp://purl.org/redcol/resource_type/TM
dc.type.versioninfo:eu-repo/semantics/acceptedVersion
dcterms.audience.professionaldevelopmentEstudiantes
dcterms.audience.professionaldevelopmentInvestigadores
dcterms.audience.professionaldevelopmentMaestros
dcterms.audience.professionaldevelopmentPúblico general
oaire.accessrightshttp://purl.org/coar/access_right/c_abf2

Archivos

Bloque original

Mostrando 1 - 1 de 1
Cargando...
Miniatura
Nombre:
1026286565_2025.pdf
Tamaño:
6.65 MB
Formato:
Adobe Portable Document Format
Descripción:
Tesis de Maestría en Ciencias - Estadística

Bloque de licencias

Mostrando 1 - 1 de 1
Cargando...
Miniatura
Nombre:
license.txt
Tamaño:
5.74 KB
Formato:
Item-specific license agreed upon to submission
Descripción: