Time Series Prediction: HMMs with TAN and Bayesian Network Observation Structures

Abstract

This thesis addresses key challenges in time series classification, focusing on enhancing predictive ac curacy through innovative modeling techniques. First, we introduce TAN-HMM, an extension of the traditional Hidden Markov Model (HMM) that incorporates Tree-Augmented Naive Bayes (TAN) to ac count for correlated features, significantly improving classification performance on complex datasets like MSRC-12. Next, we propose the Bayesian Network Hidden Markov Model (BN-HMM), which com bines the temporal dynamics of HMMs with the structural flexibility of Bayesian Networks, achieving superior accuracy and feature relationship discovery. Finally, we tackle the problem of robust early warn ing signals for disease outbreaks, utilizing cutting-edge deep learning models to predict emerging disease behavior from simulated and real-world noisy datasets. Together, these contributions push the boundaries of time series classification and offer practical solutions for real-world applications, from human activity recognition to disease outbreak prediction.