Bayesian Testing for Context Tree Equality 

Victor Freguglia Souza, IMECC-UNICAMP

Abstract:

Context trees form the foundation of various statistical methods, particularly in information compression. While estimation techniques exist to infer the context tree of a given dataset, different datasets often lead to distinct models. When dealing with multiple, potentially small datasets, storing separate models for each can be inefficient, especially as model complexity increases. In such cases, identifying groups of datasets that share a common context tree can reduce storage requirements and enhance interpretability. We propose a Bayesian method to test whether two datasets share the same context tree probability model, providing a principled approach to model selection and comparison.