Conditional probability table limit-based quantization for Bayesian networks: model quality, data fidelity and structure score

Rafael Rodrigues Mendes Ribeiro

doi:10.1007/s10489-023-05153-8

2024-04-03

Rafael Rodrigues Mendes Ribeiro *

Jordão Natal

Cassio P. de Campos

Carlos Dias Maciel

DOI：10.1007/s10489-023-05153-8

原文链接

原文求助

摘要

En 中文

Bayesian Networks (BN) are robust probabilistic graphical models mainly used with discrete random variables requiring discretization and quantization of continuous data. Quantization is known to affect model accuracy, speed and interpretability, and there are various quantization methods and performance comparisons proposed in literature. Therefore, this paper introduces a novel approach called CPT limit-based quantization (CLBQ) aimed to address the trade-off among model quality, data fidelity and structure score. CLBQ sets CPT size limitation based on how large the dataset is so as to optimize the balance between the structure score of BNs and mean squared error. For such a purpose, a range of quantization values for each variable was evaluated and a Pareto set was designed considering structure score and mean squared error (MSE). A quantization value was selected from the Pareto set in order to balance MSE and structure score, and the method's effectiveness was tested using different datasets, such as discrete variables with added noise, continuous variables and real continuous data. In all tests, CLBQ was compared to another quantization method known as Dynamic Discretization. Moreover, this study assesses the suitability of CLBQ for the search and score of BN structure learning, in addition to examining the landscape of BN structures while varying dataset sizes and confirming its consistency. It was sought to find the expected structure location through a landscape analysis and optimal BNs on it so as to confirm whether the expected results were actually achieved in the search and score of BN structure learning. Results demonstrate that CLBQ is quite capable of striking a balance between model quality, data fidelity and structure score, in addition to evidencing its potential application in the search and score of BN structure learning, thus further research should explore different structure scores and quantization methods through CLBQ. Furthermore, its code and used datasets have all been made available.

Keyword:

Quantization

Bayesian network

BIC

CPT