Date on Master's Thesis/Doctoral Dissertation

12-2014

Document Type

Doctoral Dissertation

Degree Name

Ph. D.

Department

Computer Engineering and Computer Science

Committee Chair

Frigui, Hichem

Committee Co-Chair (if applicable)

Rouchka, Eric C.

Committee Member

Yampolskiy, Roman Vladimirovich

Committee Member

Inanc, Tamer

Committee Member

Zeng, Wei-Bin

Abstract

For complex classification systems, data are gathered from various sources and potentially have different representations. Thus, data may have large intra-class variations. In fact, modeling each data class with a single model might lead to poor generalization. The classification error can be more severe for temporal data where each sample is represented by a sequence of observations. Thus, there is a need for building a classification system that takes into account the variations within each class in the data. This dissertation introduces an ensemble learning method for temporal data that uses a mixture of Hidden Markov Model (HMM) classifiers. We hypothesize that the data are generated by K models, each of which reacts a particular trend in the data. Model identification could be achieved through clustering in the feature space or in the parameters space. However, this approach is inappropriate in the context of sequential data. The proposed approach is based on clustering in the log-likelihood space, and has two main steps. First, one HMM is fit to each of the N individual sequences. For each fitted model, we evaluate the log-likelihood of each sequence. This will result in an N-by-N log-likelihood distance matrix that will be partitioned into K groups using a relational clustering algorithm. In the second step, we learn the parameters of one HMM per group. We propose using and optimizing various training approaches for the different K groups depending on their size and homogeneity. In particular, we investigate the maximum likelihood (ML), the minimum classification error (MCE) based discriminative, and the Variational Bayesian (VB) training approaches. Finally, to test a new sequence, its likelihood is computed in all the models and a final confidence value is assigned by combining the multiple models outputs using a decision level fusion method such as an artificial neural network or a hierarchical mixture of experts. Our approach was evaluated on two real-world applications: (1) identification of Cardio-Pulmonary Resuscitation (CPR) scenes in video simulating medical crises; and (2) landmine detection using Ground Penetrating Radar (GPR). Results on both applications show that the proposed method can identify meaningful and coherent HMM mixture components that describe different properties of the data. Each HMM mixture component models a group of data that share common attributes. The results indicate that the proposed method outperforms the baseline HMM that uses one model for each class in the data.

Share

COinS