• Journal of the Korean Statistical Society
• /
• 제31권1호
• /
• pp.63-75
• /
• 2002

Most attempts at Bayesian analysis of neural networks involve hierarchical modeling. We believe that similar results can be obtained with simpler models that require less computational effort, as long as appropriate restrictions are placed on parameters in order to ensure propriety of posterior distributions. In particular, we adopt a model first introduced by Lee (1999) that utilizes an improper prior for all parameters. Straightforward Gibbs sampling is possible, with the exception of the bias parameters, which are embedded in nonlinear sigmoidal functions. In addition to the problems posed by nonlinearity, direct sampling from the posterior distributions of the bias parameters is compounded due to the duplication of hidden nodes, which is a source of multimodality. In this regard, we focus on sampling from the marginal posterior distribution of the bias parameters with Markov chain Monte Carlo methods that combine traditional Metropolis sampling with a slice sampler described by Neal (1997, 2001). The methods are illustrated with data examples that are largely confined to the analysis of nonparametric regression models.

• Journal of the Korean Data and Information Science Society
• /
• 제27권1호
• /
• pp.245-254
• /
• 2016

Many surveys provide categorical data and there may be one or more missing categories. We describe a nonignorable nonresponse model for the analysis of two-way contingency tables from small areas. There are both item and unit nonresponse. One approach to analyze these data is to construct several tables corresponding to missing categories. We describe a hierarchical Bayesian model to analyze two-way categorical data from different areas. This allows a "borrowing of strength" of the data from larger areas to improve the reliability in the estimates of the model parameters corresponding to the small areas. Also we use a nonignorable nonresponse model with Bayesian uncertainty analysis by placing priors in nonidentifiable parameters instead of a sensitivity analysis for nonidentifiable parameters. We use the griddy Gibbs sampler to fit our models and compute DIC and BPP for model diagnostics. We illustrate our method using data from NHANES III data on thirteen states to obtain the finite population proportions.

• Journal of the Korean Data and Information Science Society
• /
• 제26권6호
• /
• pp.1547-1555
• /
• 2015

We study the problem of nonignorable nonresponse in a two-way contingency table and there may be one or two missing categories. We describe a nonignorable nonresponse model for the analysis of two-way categorical table. One approach to analyze these data is to construct several tables (one complete and the others incomplete). There are nonidentifiable parameters in incomplete tables. We describe a hierarchical Bayesian model to analyze two-way categorical data. We use a nonignorable nonresponse model with Bayesian uncertainty analysis by placing priors in nonidentifiable parameters instead of a sensitivity analysis for nonidentifiable parameters. To reduce the effects of nonidentifiable parameters, we project the parameters to a lower dimensional space and we allow the reduced set of parameters to share a common distribution. We use the griddy Gibbs sampler to fit our models and compute DIC and BPP for model diagnostics. We illustrate our method using data from NHANES III data to obtain the finite population proportions.

• Journal of the Korean Data and Information Science Society
• /
• 제23권3호
• /
• pp.587-593
• /
• 2012

We study Bayesian estimates for finite population proportions in multinomial problems. To do this, we consider a three-stage hierarchical Bayesian model. For prior, we use Dirichlet density to model each cell probability in each cluster. Our method does not require complicated computation such as Metropolis-Hastings algorithm to draw samples from each density of parameters. We draw samples using Gibbs sampler with grid method. We apply this algorithm to a couple of simulation data under three scenarios and we estimate the finite population proportions using two kinds of approaches We compare results with the point estimates of finite population proportions and their standard deviations. Finally, we check the consistency of computation using differen samples drawn from distinct iterates.

• Journal of the Korean Data and Information Science Society
• /
• 제28권6호
• /
• pp.1547-1555
• /
• 2017

This paper studies Bayesian test of homogeneity in contingency tables made by discretizing a continuous variable. Sometimes when we are considering events of interest in small area setup, we can think of discretization approaches about the continuous variable. If we properly discretize the continuous variable, we can find invisible relationships between areas (groups) and a continuous variable of interest. The proper discretization of the continuous variable can support the alternative hypothesis of the homogeneity test in contingency tables even if the null hypothesis was not rejected through k-sample tests involving one-way ANOVA. In other words, the proportions of variables with a particular level can vary from group to group by the discretization. If we discretize the the continuous variable, it can be treated as an analysis of the contingency table. In this case, the chi-squared test is the most commonly employed method. However, further discretization gives rise to more cells in the table. As a result, the count in the cells becomes smaller and the accuracy of the test becomes lower. To prevent this, we can consider the Bayesian approach and apply it to the setup of the homogeneity test.

• 한국환경과학회지
• /
• 제27권7호
• /
• pp.499-507
• /
• 2018

Agricultural meteorological information is an important resource that affects farmers' income, food security, and agricultural conditions. Thus, such data are used in various fields that are responsible for planning, enforcing, and evaluating agricultural policies. The meteorological information obtained from automatic weather observation systems operated by rural development agencies contains missing values owing to temporary mechanical or communication deficiencies. It is known that missing values lead to reduction in the reliability and validity of the model. In this study, the hierarchical Bayesian spatio-temporal model suggests replacements for missing values because the meteorological information includes spatio-temporal correlation. The prior distribution is very important in the Bayesian approach. However, we found a problem where the spatial decay parameter was not converged through the trace plot. A suitable spatial decay parameter, estimated on the bias of root-mean-square error (RMSE), which was determined to be the difference between the predicted and observed values. The latitude, longitude, and altitude were considered as covariates. The estimated spatial decay parameters were 0.041 and 0.039, for the spatio-temporal model with latitude and longitude and for latitude, longitude, and altitude, respectively. The posterior distributions were stable after the spatial decay parameter was fixed. root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), and bias were calculated for model validation. Finally, the missing values were generated using the independent Gaussian process model.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• 제12권12호
• /
• pp.5819-5840
• /
• 2018

Considering the topology of hierarchical tree structure, each cluster in WSNs is faced with various attacks launched by malicious nodes, which include network eavesdropping, channel interference and data tampering. The existing intrusion detection algorithm does not take into consideration the resource constraints of cluster heads and sensor nodes. Due to application requirements, sensor nodes in WSNs are deployed with approximately uncorrelated security weights. In our study, a novel and versatile intrusion detection system (IDS) for the optimal defense strategy is primarily introduced. Given the flexibility that wireless communication provides, it is unreasonable to expect malicious nodes will demonstrate a fixed behavior over time. Instead, malicious nodes can dynamically update the attack strategy in response to the IDS in each game stage. Thus, a multi-stage intrusion detection game (MIDG) based on Bayesian rules is proposed. In order to formulate the solution of MIDG, an in-depth analysis on the Bayesian equilibrium is performed iteratively. Depending on the MIDG theoretical analysis, the optimal behaviors of rational attackers and defenders are derived and calculated accurately. The numerical experimental results validate the effectiveness and robustness of the proposed scheme.

• Journal of the Korean Statistical Society
• /
• 제30권2호
• /
• pp.179-192
• /
• 2001

When a statistical model has a hierarchical structure such as multilayer perceptrons in neural networks or Gaussian mixture density representation, the model includes distribution with unidentifiable parameters when the structure becomes redundant. Since the exact structure is unknown, we need to carry out statistical estimation or learning of parameters in such a model. From the geometrical point of view, distributions specified by unidentifiable parameters become a singular point in the parameter space. The problem has been remarked in many statistical models, and strange behaviors of the likelihood ratio statistics, when the null hypothesis is at a singular point, have been analyzed so far. The present paper studies asymptotic behaviors of the maximum likelihood estimator and the Bayesian predictive estimator, by using a simple cone model, and show that they are completely different from regular statistical models where the Cramer-Rao paradigm holds. At singularities, the Fisher information metric degenerates, implying that the cramer-Rao paradigm does no more hold, and that he classical model selection theory such as AIC and MDL cannot be applied. This paper is a first step to establish a new theory for analyzing the accuracy of estimation or learning at around singularities.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• 제11권12호
• /
• pp.5905-5926
• /
• 2017

Sentiments can profoundly affect individual behavior as well as decision-making. Confronted with the ever-increasing amount of review information available online, it is desirable to provide an effective sentiment model to both detect and organize the available information to improve understanding, and to present the information in a more constructive way for consumers. This study developed a unified phrase-based topic and sentiment detection model, combined with a tracking model using incremental hierarchical dirichlet allocation (PTSM_IHDP). This model was proposed to discover the evolutionary trend of topic-based sentiments from online reviews. PTSM_IHDP model firstly assumed that each review document has been composed by a series of independent phrases, which can be represented as both topic information and sentiment information. PTSM_IHDP model secondly depended on an improved time-dependency non-parametric Bayesian model, integrating incremental hierarchical dirichlet allocation, to estimate the optimal number of topics by incrementally building an up-to-date model. To evaluate the effectiveness of our model, we tested our model on a collected dataset, and compared the result with the predictions of traditional models. The results demonstrate the effectiveness and advantages of our model compared to several state-of-the-art methods.

• Management Science and Financial Engineering
• /
• 제13권1호
• /
• pp.73-91
• /
• 2007

We investigate factors that influence the choice of high-share brands(HSBs) vs. low-share brands(LSBs) among various product and consumer characteristics related to brand-share perceptions. Specifically, using 8 product categories varying in terms of purchase decision involvement, we show how the influencing factors vary across the categories. At the general level that cover all the 8 categories, our hierarchical Bayesian regressions analysis shows that factors that favor high-share brands are purchase decision involvement, search goods, experience goods, price-quality relationship, positive network externalities, and price-prestige beliefs. Conversely, consumers who value variety seeking and need for uniqueness favor low-share brands. The effects of these factors, however, vary across product categories. The identification of these characteristics can help brand managers establish a more effective brand-share strategy in such areas as setting an optimal market share goal, extending a brand, and developing ad copy. Furthermore, our consumer segmentation analysis demonstrates the general market has two distinct segments - (1) a segment composed of HSB buyers(86%) and (2) a segment composed of LSB buyers(14%). The two segments are also shown to have different significant factors that explain their brand choice. Our segmentation analysis can help marketers establish a marketing strategy that targets a specific segment of interest.