ApplyingtheHiddenMarkovModelMethodologyforUnsupe.ppt

Applying the Hidden Markov Model Methodology for Unsupervised Learning of Temporal Data Cen Li1 and Gautam Biswas2 1Department of Computer Science Middle Tennessee State University 2Department of EECS Vanderbilt University Nashville, TN 37235. USA. biswas@ biswas@/~biswas June 21, 2001 Problem Description Most real world systems are dynamic, e.g., Physical plants and Engineering systems Human Physiology Economic systems Systems are complex: hard to understand, model, and analyze But our ability to collect data on these systems has increased tremendously Task: Use data to automatically build models, extend incomplete models, and verify and validate existing models using the data available. Why models? Formal, abstract representation of phenomena or process Enables systematic analysis and prediction Our goal: Build models, i.e., create structure from data using exploratory techniques Challenge: Systematic and useful clustering algorithms for temporal data Outline of Talk Example Problem Related work on temporal data clustering Motivation for using Hidden Markov Model (HMM) representation Bayesian HMM clustering methodology Experimental results Synthetic Data Real world ecological data Conclusions and future work Example of Dynamic System Problem Description Unsupervised Classification (Clustering) Given data objects described by multiple temporal features: (1) to assign category information to individual objects by objectively partitioning the objects into homogeneous groups such that the within group object similarity and the between group object dissimilarity are maximized, (2) to form succinct description for each category derived. Related Work on Temporal Data Clustering Motivation for using HMM Representation What are our modeling objectives ? Why do we choose the HMM representation ? The hidden states valid stages of a dynamic process Direct probabilistic links probabilistic transitions among different stages