Mutual Information and Perplexity based Clustering of Dialogue Information for Dynamic Adaptation of Language Models

We present two approaches to cluster dialogue-based information obtained by the speech understanding module and the dialogue manager of a spoken dialogue system. The purpose is to estimate a language model related to each cluster, and use them to dynamically modify the model of the speech recognizer at each dialogue turn. In the first approach we build the cluster tree using local decisions based on a Maximum Normalized Mutual Information criterion. In the second one we take global decisions, based on the optimization of the global perplexity of the combination of the cluster-related LMs. Our experiments show a relative reduction of the word error rate of 15.17%, which helps to improve the performance of the understanding and the dialogue manager modules.

Mutual Information and Perplexity based Clustering of Dialogue Information for Dynamic Adaptation of Language Models

We present two approaches to cluster dialogue-based information obtained by the speech understanding module and the dialogue manager of a spoken dialogue system. The purpose is to estimate a language model related to each cluster, and use them to dynamically modify the model of the speech recognizer at each dialogue turn. In the first approach we build the cluster tree using local decisions based on a Maximum Normalized Mutual Information criterion. In the second one we take global decisions, based on the optimization of the global perplexity of the combination of the cluster-related LMs. Our experiments show a relative reduction of the word error rate of 15.17%, which helps to improve the performance of the understanding and the dialogue manager modules.

Bayes factor based speaker clustering for speaker diarization

This paper proposes the use of the Bayes Factor to replace the Bayesian Information Criterion (BIC) as a criterion for speaker clustering within a speaker diarization system. The BIC is one of the most popular decision criteria used in speaker diarization systems today. However, it will be shown in this paper that the BIC is only an approximation to the Bayes factor of marginal likelihoods of the data given each hypothesis. This paper uses the Bayes factor directly as a decision criterion for speaker clustering, thus removing the error introduced by the BIC approximation. Results obtained on the 2002 Rich Transcription (RT-02) Evaluation dataset show an improved clustering performance, leading to a 14.7% relative improvement in the overall Diarization Error Rate (DER) compared to the baseline system.

Addressing issues in sparseness, ecological bias and formulation of the adjacency matrix in Bayesian spatio-temporal analysis of disease counts

The main objective of this PhD was to further develop Bayesian spatio-temporal models (specifically the Conditional Autoregressive (CAR) class of models), for the analysis of sparse disease outcomes such as birth defects. The motivation for the thesis arose from problems encountered when analyzing a large birth defect registry in New South Wales. The specific components and related research objectives of the thesis were developed from gaps in the literature on current formulations of the CAR model, and health service planning requirements. Data from a large probabilistically-linked database from 1990 to 2004, consisting of fields from two separate registries: the Birth Defect Registry (BDR) and Midwives Data Collection (MDC) were used in the analyses in this thesis. The main objective was split into smaller goals. The first goal was to determine how the specification of the neighbourhood weight matrix will affect the smoothing properties of the CAR model, and this is the focus of chapter 6. Secondly, I hoped to evaluate the usefulness of incorporating a zero-inflated Poisson (ZIP) component as well as a shared-component model in terms of modeling a sparse outcome, and this is carried out in chapter 7. The third goal was to identify optimal sampling and sample size schemes designed to select individual level data for a hybrid ecological spatial model, and this is done in chapter 8. Finally, I wanted to put together the earlier improvements to the CAR model, and along with demographic projections, provide forecasts for birth defects at the SLA level. Chapter 9 describes how this is done. For the first objective, I examined a series of neighbourhood weight matrices, and showed how smoothing the relative risk estimates according to similarity by an important covariate (i.e. maternal age) helped improve the model’s ability to recover the underlying risk, as compared to the traditional adjacency (specifically the Queen) method of applying weights. Next, to address the sparseness and excess zeros commonly encountered in the analysis of rare outcomes such as birth defects, I compared a few models, including an extension of the usual Poisson model to encompass excess zeros in the data. This was achieved via a mixture model, which also encompassed the shared component model to improve on the estimation of sparse counts through borrowing strength across a shared component (e.g. latent risk factor/s) with the referent outcome (caesarean section was used in this example). Using the Deviance Information Criteria (DIC), I showed how the proposed model performed better than the usual models, but only when both outcomes shared a strong spatial correlation. The next objective involved identifying the optimal sampling and sample size strategy for incorporating individual-level data with areal covariates in a hybrid study design. I performed extensive simulation studies, evaluating thirteen different sampling schemes along with variations in sample size. This was done in the context of an ecological regression model that incorporated spatial correlation in the outcomes, as well as accommodating both individual and areal measures of covariates. Using the Average Mean Squared Error (AMSE), I showed how a simple random sample of 20% of the SLAs, followed by selecting all cases in the SLAs chosen, along with an equal number of controls, provided the lowest AMSE. The final objective involved combining the improved spatio-temporal CAR model with population (i.e. women) forecasts, to provide 30-year annual estimates of birth defects at the Statistical Local Area (SLA) level in New South Wales, Australia. The projections were illustrated using sixteen different SLAs, representing the various areal measures of socio-economic status and remoteness. A sensitivity analysis of the assumptions used in the projection was also undertaken. By the end of the thesis, I will show how challenges in the spatial analysis of rare diseases such as birth defects can be addressed, by specifically formulating the neighbourhood weight matrix to smooth according to a key covariate (i.e. maternal age), incorporating a ZIP component to model excess zeros in outcomes and borrowing strength from a referent outcome (i.e. caesarean counts). An efficient strategy to sample individual-level data and sample size considerations for rare disease will also be presented. Finally, projections in birth defect categories at the SLA level will be made.

On the nature of over-dispersion in motor vehicle crash prediction models

Statistical modeling of traffic crashes has been of interest to researchers for decades. Over the most recent decade many crash models have accounted for extra-variation in crash counts—variation over and above that accounted for by the Poisson density. The extra-variation – or dispersion – is theorized to capture unaccounted for variation in crashes across sites. The majority of studies have assumed fixed dispersion parameters in over-dispersed crash models—tantamount to assuming that unaccounted for variation is proportional to the expected crash count. Miaou and Lord [Miaou, S.P., Lord, D., 2003. Modeling traffic crash-flow relationships for intersections: dispersion parameter, functional form, and Bayes versus empirical Bayes methods. Transport. Res. Rec. 1840, 31–40] challenged the fixed dispersion parameter assumption, and examined various dispersion parameter relationships when modeling urban signalized intersection accidents in Toronto. They suggested that further work is needed to determine the appropriateness of the findings for rural as well as other intersection types, to corroborate their findings, and to explore alternative dispersion functions. This study builds upon the work of Miaou and Lord, with exploration of additional dispersion functions, the use of an independent data set, and presents an opportunity to corroborate their findings. Data from Georgia are used in this study. A Bayesian modeling approach with non-informative priors is adopted, using sampling-based estimation via Markov Chain Monte Carlo (MCMC) and the Gibbs sampler. A total of eight model specifications were developed; four of them employed traffic flows as explanatory factors in mean structure while the remainder of them included geometric factors in addition to major and minor road traffic flows. The models were compared and contrasted using the significance of coefficients, standard deviance, chi-square goodness-of-fit, and deviance information criteria (DIC) statistics. The findings indicate that the modeling of the dispersion parameter, which essentially explains the extra-variance structure, depends greatly on how the mean structure is modeled. In the presence of a well-defined mean function, the extra-variance structure generally becomes insignificant, i.e. the variance structure is a simple function of the mean. It appears that extra-variation is a function of covariates when the mean structure (expected crash count) is poorly specified and suffers from omitted variables. In contrast, when sufficient explanatory variables are used to model the mean (expected crash count), extra-Poisson variation is not significantly related to these variables. If these results are generalizable, they suggest that model specification may be improved by testing extra-variation functions for significance. They also suggest that known influences of expected crash counts are likely to be different than factors that might help to explain unaccounted for variation in crashes across sites

Improving the performance of model-order selection criteria by partial-model selection search

The traditional searching method for model-order selection in linear regression is a nested full-parameters-set searching procedure over the desired orders, which we call full-model order selection. On the other hand, a method for model-selection searches for the best sub-model within each order. In this paper, we propose using the model-selection searching method for model-order selection, which we call partial-model order selection. We show by simulations that the proposed searching method gives better accuracies than the traditional one, especially for low signal-to-noise ratios over a wide range of model-order selection criteria (both information theoretic based and bootstrap-based). Also, we show that for some models the performance of the bootstrap-based criterion improves significantly by using the proposed partial-model selection searching method. Index Terms— Model order estimation, model selection, information theoretic criteria, bootstrap 1. INTRODUCTION Several model-order selection criteria can be applied to find the optimal order. Some of the more commonly used information theoretic-based procedures include Akaike’s information criterion (AIC) [1], corrected Akaike (AICc) [2], minimum description length (MDL) [3], normalized maximum likelihood (NML) [4], Hannan-Quinn criterion (HQC) [5], conditional model-order estimation (CME) [6], and the efficient detection criterion (EDC) [7]. From a practical point of view, it is difficult to decide which model order selection criterion to use. Many of them perform reasonably well when the signal-to-noise ratio (SNR) is high. The discrepancies in their performance, however, become more evident when the SNR is low. In those situations, the performance of the given technique is not only determined by the model structure (say a polynomial trend versus a Fourier series) but, more importantly, by the relative values of the parameters within the model. This makes the comparison between the model-order selection algorithms difficult as within the same model with a given order one could find an example for which one of the methods performs favourably well or fails [6, 8]. Our aim is to improve the performance of the model order selection criteria in cases where the SNR is low by considering a model-selection searching procedure that takes into account not only the full-model order search but also a partial model order search within the given model order. Understandably, the improvement in the performance of the model order estimation is at the expense of additional computational complexity.

Bayesian methodology for genetics of complex diseases

Genetic research of complex diseases is a challenging, but exciting, area of research. The early development of the research was limited, however, until the completion of the Human Genome and HapMap projects, along with the reduction in the cost of genotyping, which paves the way for understanding the genetic composition of complex diseases. In this thesis, we focus on the statistical methods for two aspects of genetic research: phenotype definition for diseases with complex etiology and methods for identifying potentially associated Single Nucleotide Polymorphisms (SNPs) and SNP-SNP interactions. With regard to phenotype definition for diseases with complex etiology, we firstly investigated the effects of different statistical phenotyping approaches on the subsequent analysis. In light of the findings, and the difficulties in validating the estimated phenotype, we proposed two different methods for reconciling phenotypes of different models using Bayesian model averaging as a coherent mechanism for accounting for model uncertainty. In the second part of the thesis, the focus is turned to the methods for identifying associated SNPs and SNP interactions. We review the use of Bayesian logistic regression with variable selection for SNP identification and extended the model for detecting the interaction effects for population based case-control studies. In this part of study, we also develop a machine learning algorithm to cope with the large scale data analysis, namely modified Logic Regression with Genetic Program (MLR-GEP), which is then compared with the Bayesian model, Random Forests and other variants of logic regression.

An economic evaluation of the Safe Motherhood programme in Guangxi, China

Maternal and infant mortality is a global health issue with a significant social and economic impact. Each year, over half a million women worldwide die due to complications related to pregnancy or childbirth, four million infants die in the first 28 days of life, and eight million infants die in the first year. Ninety-nine percent of maternal and infant deaths are in developing countries. Reducing maternal and infant mortality is among the key international development goals. In China, the national maternal mortality ratio and infant mortality rate were reduced greatly in the past two decades, yet a large discrepancy remains between urban and rural areas. To address this problem, a large-scale Safe Motherhood Programme was initiated in 2000. The programme was implemented in Guangxi in 2003. Interventions in the programme included both demand-side and supply side-interventions focusing on increasing health service use and improving birth outcomes. Little is known about the effects and economic outcomes of the Safe Motherhood Programme in Guangxi, although it has been implemented for seven years. The aim of this research is to estimate the effectiveness and cost-effectiveness of the interventions in the Safe Motherhood Programme in Guangxi, China. The objectives of this research include: 1. To evaluate whether the changes of health service use and birth outcomes are associated with the interventions in the Safe Motherhood Programme. 2. To estimate the cost-effectiveness of the interventions in the Safe Motherhood Programme and quantify the uncertainty surrounding the decision. 3. To assess the expected value of perfect information associated with both the whole decision and individual parameters, and interpret the findings to inform priority setting in further research and policy making in this area. A quasi-experimental study design was used in this research to assess the effectiveness of the programme in increasing health service use and improving birth outcomes. The study subjects were 51 intervention counties and 30 control counties. Data on the health service use, birth outcomes and socio-economic factors from 2001 to 2007 were collected from the programme database and statistical yearbooks. Based on the profile plots of the data, general linear mixed models were used to evaluate the effectiveness of the programme while controlling for the effects of baseline levels of the response variables, change of socio-economic factors over time and correlations among repeated measurements from the same county. Redundant multicollinear variables were deleted from the mixed model using the results of the multicollinearity diagnoses. For each response variable, the best covariance structure was selected from 15 alternatives according to the fit statistics including Akaike information criterion, Finite-population corrected Akaike information criterion, and Schwarz.s Bayesian information criterion. Residual diagnostics were used to validate the model assumptions. Statistical inferences were made to show the effect of the programme on health service use and birth outcomes. A decision analytic model was developed to evaluate the cost-effectiveness of the programme, quantify the decision uncertainty, and estimate the expected value of perfect information associated with the decision. The model was used to describe the transitions between health states for women and infants and reflect the change of both costs and health benefits associated with implementing the programme. Result gained from the mixed models and other relevant evidence identified were synthesised appropriately to inform the input parameters of the model. Incremental cost-effectiveness ratios of the programme were calculated for the two groups of intervention counties over time. Uncertainty surrounding the parameters was dealt with using probabilistic sensitivity analysis, and uncertainty relating to model assumptions was handled using scenario analysis. Finally the expected value of perfect information for both the whole model and individual parameters in the model were estimated to inform priority setting in further research in this area.The annual change rates of the antenatal care rate and the institutionalised delivery rate were improved significantly in the intervention counties after the programme was implemented. Significant improvements were also found in the annual change rates of the maternal mortality ratio, the infant mortality rate, the incidence rate of neonatal tetanus and the mortality rate of neonatal tetanus in the intervention counties after the implementation of the programme. The annual change rate of the neonatal mortality rate was also improved, although the improvement was only close to statistical significance. The influences of the socio-economic factors on the health service use indicators and birth outcomes were identified. The rural income per capita had a significant positive impact on the health service use indicators, and a significant negative impact on the birth outcomes. The number of beds in healthcare institutions per 1,000 population and the number of rural telephone subscribers per 1,000 were found to be positively significantly related to the institutionalised delivery rate. The length of highway per square kilometre negatively influenced the maternal mortality ratio. The percentage of employed persons in the primary industry had a significant negative impact on the institutionalised delivery rate, and a significant positive impact on the infant mortality rate and neonatal mortality rate. The incremental costs of implementing the programme over the existing practice were US $11.1 million from the societal perspective, and US $13.8 million from the perspective of the Ministry of Health. Overall, 28,711 life years were generated by the programme, producing an overall incremental cost-effectiveness ratio of US $386 from the societal perspective, and US $480 from the perspective of the Ministry of Health, both of which were below the threshold willingness-to-pay ratio of US $675. The expected net monetary benefit generated by the programme was US $8.3 million from the societal perspective, and US $5.5 million from the perspective of the Ministry of Health. The overall probability that the programme was cost-effective was 0.93 and 0.89 from the two perspectives, respectively. The incremental cost-effectiveness ratio of the programme was insensitive to the different estimates of the three parameters relating to the model assumptions. Further research could be conducted to reduce the uncertainty surrounding the decision, in which the upper limit of investment was US $0.6 million from the societal perspective, and US $1.3 million from the perspective of the Ministry of Health. It is also worthwhile to get a more precise estimate of the improvement of infant mortality rate. The population expected value of perfect information for individual parameters associated with this parameter was US $0.99 million from the societal perspective, and US $1.14 million from the perspective of the Ministry of Health. The findings from this study have shown that the interventions in the Safe Motherhood Programme were both effective and cost-effective in increasing health service use and improving birth outcomes in rural areas of Guangxi, China. Therefore, the programme represents a good public health investment and should be adopted and further expanded to an even broader area if possible. This research provides economic evidence to inform efficient decision making in improving maternal and infant health in developing countries.

The effect of various temperature indicators on different mortality categories in a subtropical city of Brisbane, Australia.

BACKGROUND: The relationship between temperature and mortality has been explored for decades and many temperature indicators have been applied separately. However, few data are available to show how the effects of different temperature indicators on different mortality categories, particularly in a typical subtropical climate. OBJECTIVE: To assess the associations between various temperature indicators and different mortality categories in Brisbane, Australia during 1996-2004. METHODS: We applied two methods to assess the threshold and temperature indicator for each age and death groups: mean temperature and the threshold assessed from all cause mortality was used for all mortality categories; the specific temperature indicator and the threshold for each mortality category were identified separately according to the minimisation of AIC. We conducted polynomial distributed lag non-linear model to identify effect estimates in mortality with one degree of temperature increase (or decrease) above (or below) the threshold on current days and lagged effects using both methods. RESULTS: Akaike's Information Criterion was minimized when mean temperature was used for all non-external deaths and deaths from 75 to 84 years; when minimum temperature was used for deaths from 0 to 64 years, 65-74 years, ≥ 85 years, and from the respiratory diseases; when maximum temperature was used for deaths from cardiovascular diseases. The effect estimates using certain temperature indicators were similar as mean temperature both for current day and lag effects. CONCLUSION: Different age groups and death categories were sensitive to different temperature indicators. However, the effect estimates from certain temperature indicators did not significantly differ from those of mean temperature.

Automatic Audio Segmentation Using the Generalized Likelihood Ratio

This paper presents a novel technique for segmenting an audio stream into homogeneous regions according to speaker identities, background noise, music, environmental and channel conditions. Audio segmentation is useful in audio diarization systems, which aim to annotate an input audio stream with information that attributes temporal regions of the audio into their specific sources. The segmentation method introduced in this paper is performed using the Generalized Likelihood Ratio (GLR), computed between two adjacent sliding windows over preprocessed speech. This approach is inspired by the popular segmentation method proposed by the pioneering work of Chen and Gopalakrishnan, using the Bayesian Information Criterion (BIC) with an expanding search window. This paper will aim to identify and address the shortcomings associated with such an approach. The result obtained by the proposed segmentation strategy is evaluated on the 2002 Rich Transcription (RT-02) Evaluation dataset, and a miss rate of 19.47% and a false alarm rate of 16.94% is achieved at the optimal threshold.

Displacement motion prediction of a landing deck for recovery operations of rotary UAVs

This paper proposes a practical prediction procedure for vertical displacement of a Rotarywing Unmanned Aerial Vehicle (RUAV) landing deck in the presence of stochastic sea state disturbances. A proper time series model tending to capture characteristics of the dynamic relationship between an observer and a landing deck is constructed, with model orders determined by a novel principle based on Bayes Information Criterion (BIC) and coefficients identified using the Forgetting Factor Recursive Least Square (FFRLS) method. In addition, a fast-converging online multi-step predictor is developed, which can be implemented more rapidly than the Auto-Regressive (AR) predictor as it requires less memory allocations when updating coefficients. Simulation results demonstrate that the proposed prediction approach exhibits satisfactory prediction performance, making it suitable for integration into ship-helicopter approach and landing guidance systems in consideration of computational capacity of the flight computer.

Modelling survival data to account for model uncertainty: a single model or model averaging?

This study considered the problem of predicting survival, based on three alternative models: a single Weibull, a mixture of Weibulls and a cure model. Instead of the common procedure of choosing a single “best” model, where “best” is defined in terms of goodness of fit to the data, a Bayesian model averaging (BMA) approach was adopted to account for model uncertainty. This was illustrated using a case study in which the aim was the description of lymphoma cancer survival with covariates given by phenotypes and gene expression. The results of this study indicate that if the sample size is sufficiently large, one of the three models emerge as having highest probability given the data, as indicated by the goodness of fit measure; the Bayesian information criterion (BIC). However, when the sample size was reduced, no single model was revealed as “best”, suggesting that a BMA approach would be appropriate. Although a BMA approach can compromise on goodness of fit to the data (when compared to the true model), it can provide robust predictions and facilitate more detailed investigation of the relationships between gene expression and patient survival. Keywords: Bayesian modelling; Bayesian model averaging; Cure model; Markov Chain Monte Carlo; Mixture model; Survival analysis; Weibull distribution

Changes in healthcare-associated Staphylococcus aureus bloodstream infections after the introduction of a national hand hygiene initiative

Background. Interventions that prevent healthcare-associated infection should lead to fewer deaths and shorter hospital stays. Cleaning hands (with soap or alcohol) is an effective way to prevent the transmission of organisms, but rates of compliance with hand hygiene are sometimes disappointingly low. The National Hand Hygiene Initiative in Australia aimed to improve hand hygiene compliance among healthcare workers, with the goal of reducing rates of healthcare-associated infection. Methods. We examined whether the introduction of the National Hand Hygiene Initiative was associated with a change in infection rates. Monthly infection rates for healthcare-associated Staphylococcus aureus bloodstream infections were examined in 38 Australian hospitals across 6 states. We used Poisson regression and examined 12 possible patterns of change, with the best fitting pattern chosen using the Akaike information criterion. Monthly bed-days were included to control for increased hospital use over time. Results. The National Hand Hygiene Initiative was associated with a reduction in infection rates in 4 of the 6 states studied. Two states showed an immediate reduction in rates of 17% and 28%, 2 states showed a linear decrease in rates of 8% and 11% per year, and 2 showed no change in infection rates. Conclusions. The intervention was associated with reduced infection rates in most states. The failure in 2 states may have been because those states already had effective initiatives before the national initiative’s introduction or because infection rates were already low and could not be further reduced.

Minimizing the effects of overfitting and collinearity in construction cost estimation : a new hybird approach

Research problem: Overfitting and collinearity problems commonly exist in current construction cost estimation applications and obstruct researchers and practitioners in achieving better modelling results. Research objective and method: A hybrid approach of Akaike information criterion (AIC) stepwise regression and principal component regression (PCR) is proposed to help solve overfitting and collinearity problems. Utilization of this approach in linear regression is validated by comparing it with other commonly used approaches. The mean square error obtained by leave-one-out cross validation (MSELOOCV) is used in model selection in deciding predictive variables.