Poon, Ka Ho. / Thesis (M.Phil.)--Chinese University of Hong Kong, 2011. / Includes bibliographical references (leaves 62-63). / Abstracts in English and Chinese. / Abstract of the thesis entitled --- p.ii / 摘要 --- p.iii / Acknowledgements --- p.iv / Chapter Section 1. --- Introduction --- p.1 / Chapter Section 2. --- Review of Useful Concepts --- p.4 / Chapter 2.1 --- Terms and Methodologies - Pattern Recognition --- p.4 / Chapter 2.1.1 --- Rolling Windows --- p.4 / Chapter 2.1.2 --- Smoothing Function - Kernel Regression --- p.5 / Chapter 2.1.3 --- Filtering Function ´ؤ Search for Extrema --- p.6 / Chapter 2.1.4 --- Filtering Function - The Pattern Detection Algorithm --- p.7 / Chapter 2.1.5 --- Risk-adjustment Model --- p.10 / Chapter Section 3. --- Data and Methodology --- p.12 / Chapter 3.1 --- Data --- p.12 / Chapter 3.2 --- Methodology --- p.12 / Chapter Section 4. --- Results --- p.17 / Chapter Section 5. --- Further Extension --- p.21 / Chapter Section 6. --- Discussions and Conclusion --- p.22 / APPENDIX 1 --- p.23 / References --- p.62

Traditional stochastic approaches for synthetic generation of weather variables often assume a prior functional form for the stochastic process, are often not capable of reproducing the probabilistic structure present in the data, and may not be uniformly applicable across sites. In an attempt to find a general framework for stochastic generation of weather variables, this study marks a unique departure from the traditional approaches, and ushers in the use of data-driven nonparametric techniques and demonstrates their utility. Precipitation is one of the key variables that drive hydrologic systems and hence warrants more focus . In this regard, two major aspects of precipitation modeling were considered: (I) resampling traces under the assumption of stationarity in the process, or with some treatment of the seasonality, and (2) investigations into interannual and secular trends in precipitation and their likely implications. A nonparametric seasonal wet/dry spell model was developed for the generation of daily precipitation. In this the probability density functions of interest are estimated using non parametric kernel density estimators. In the course of development of this model, various nonparametric density estimators for discrete and continuous data were reviewed, tested, and documented, which resulted in the development of a nonparametric estimator for discrete probability estimation. Variations in seasonality of precipitation as a function of latitude and topographic factors were seen through the non parametric estimation of the time-varying occurrence frequency. Nonparametric spectral analysis, performed on monthly precipitation, revealed significant interannual frequencies and coherence with known atmospheric oscillations. Consequently, a non parametric, nonhomogeneous Markov chain for modeling daily precipitation was developed that obviated the need to divide the year into seasons. Multivariate nonparametric resampling technique from the nonparametrically fitted probability density functions, which can be likened to a smoothed bootstrap approach, was developed for the simulation of other weather variables (solar radiation, maximum and minimum temperature, average dew point temperature, and average wind speed). In this technique the vector of variables on a day is generated by conditioning on the vector of these variables on the preceding day and the precipitation amount on the current day generated from the wet/dry spell model.

當我們探討的是兩組樣本的變異是否有所差異時，常見的方法有以ANOVA 為 基礎的檢定與秩檢定，傳統的秩檢定需要假設兩母體具有相同的中位數或知道 其差異。本研究採用Moses (1963) 提出的rank-like 檢定方法，此方法在處理兩組樣本的變異問題時，優點是不需要估計任何中心參數，也不需要假設母體中心參數相同，在資料偏態的情況下也表現得很穩健，我們試圖在樣本數極小的情況下對此方法作修正，將此檢定方法與以ANOVA 為基礎的檢定和秩檢定進行模擬比較，以能夠良好的控制型一誤差與檢定力作為評斷標準。由模擬的結果可得知，rank-like 檢定方法與修正後的方法在不同的分配下皆表現的穩健而修正後的方法特別適用於小樣本的情形。 / We consider the problem of detecting variability change in the two-sample case.Several classical variability tests are investigated, including the ANOVA based tests and the rank tests. Traditional two-sample rank tests assume that the location parameters for both samples are identical or of known difference. In this thesis, a modified version of the distribution-free rank-like test proposed by Moses (1963) is proposed. Moses’s test has several advantages. It does not require location parameter estimation, is applicable without assuming that location parameter are identical, and is robust for skewed data. However, Moses’s test has no power when each of the two samples has size 5 or less. The modified version of Moses’s test proposed in this thesis has some power when the sample sizes are small. Comparative simulation results are presented. According to these results, both Moses’s test and the proposed test are robust under all conditions, and the proposed test works better when the sample sizes are small.

Semiparametric maximum likelihood analysis allows inference in errors-invariables models with small loss of efficiency relative to full likelihood analysis but with significantly weakened assumptions. In addition, since no distributional assumptions are made for the nuisance parameters, the analysis more nearly parallels that for usual regression. These highly desirable features and the high degree of modelling flexibility permitted warrant the development of the approach for routine use. This thesis does so for the special cases of linear and nonlinear regression with measurement errors in one explanatory variable. A transparent and flexible computational approach is developed, the analysis is exhibited on some examples, and finite sample properties of estimates, approximate standard errors, and likelihood ratio inference are clarified with simulation. / Graduation date: 2001

In many applications of graphical models arising in computer vision, the hidden variables of interest are most naturally specified by continuous, non-Gaussian distributions. There exist inference algorithms for discrete approximations to these continuous distributions, but for the high-dimensional variables typically of interest, discrete inference becomes infeasible. Stochastic methods such as particle filters provide an appealing alternative. However, existing techniques fail to exploit the rich structure of the graphical models describing many vision problems. Drawing on ideas from regularized particle filters and belief propagation (BP), this paper develops a nonparametric belief propagation (NBP) algorithm applicable to general graphs. Each NBP iteration uses an efficient sampling procedure to update kernel-based approximations to the true, continuous likelihoods. The algorithm can accomodate an extremely broad class of potential functions, including nonparametric representations. Thus, NBP extends particle filtering methods to the more general vision problems that graphical models can describe. We apply the NBP algorithm to infer component interrelationships in a parts-based face model, allowing location and reconstruction of occluded features.

This dissertation focuses on the study of nonparametric quickest detection and its decentralized implementation in a distributed environment. Quickest detection schemes are geared toward detecting a change in the state of a data stream or a real-time process. Classical quickest detection schemes invariably assume knowledge of the pre-change and post-change distributions that may not be available in many applications. A distribution free nonparametric quickest detection procedure is presented based on a novel distance measure, referred to as the Q-Q distance calculated from the Quantile-Quantile plot. Theoretical analysis of the distance measure and detection procedure is presented to justify the proposed algorithm and provide performance guarantees. The Q-Q distance based detection procedure presents comparable performance compared to classical parametric detection procedure and better performance than other nonparametric procedures. The proposed procedure is most effective when detecting small changes. As the technology advances, distributed sensing and detection become feasible. Existing decentralized detection approaches are largely parametric. The decentralized realization of Q-Q distance based nonparametric quickest detection scheme is further studied, where data streams are simultaneously collected from multiple channels located distributively to jointly reach a detection decision. Two implementation schemes, binary quickest detection and local decision fusion, are described. Experimental results show that the proposed method has a comparable performance to the benchmark parametric cumulative sum (CUSUM) test in binary detection. Finally the dissertation concludes with a summary of the contributions to the state of the art.

Non-parametric belief propagation (NBP) is a well-known message passing method for cooperative localization in wireless networks. However, due to the over-counting problem in the networks with loops, NBP’s convergence is not guaranteed, and its estimates are typically less accurate. One solution for this problem is non-parametric generalized belief propagation based on junction tree. However, this method is intractable in large-scale networks due to the high-complexity of the junction tree formation, and the high-dimensionality of the particles. Therefore, in this article, we propose the non-parametric generalized belief propagation based on pseudo-junction tree (NGBP-PJT). The main difference comparing with the standard method is the formation of pseudo-junction tree, which represents the approximated junction tree based on thin graph. In addition, in order to decrease the number of high-dimensional particles, we use more informative importance density function, and reduce the dimensionality of the messages. As by-product, we also propose NBP based on thin graph (NBP-TG), a cheaper variant of NBP, which runs on the same graph as NGBP-PJT. According to our simulation and experimental results, NGBP-PJT method outperforms NBP and NBP-TG in terms of accuracy, computational, and communication cost in reasonably sized networks. / COOPLOC / FP7-ICT WHERE2

<p>We propose nonparametric Bayesian models for supervised dimension</p><p>reduction and regression problems. Supervised dimension reduction is</p><p>a setting where one needs to reduce the dimensionality of the</p><p>predictors or find the dimension reduction subspace and lose little</p><p>or no predictive information. Our first method retrieves the</p><p>dimension reduction subspace in the inverse regression framework by</p><p>utilizing a dependent Dirichlet process that allows for natural</p><p>clustering for the data in terms of both the response and predictor</p><p>variables. Our second method is based on ideas from the gradient</p><p>learning framework and retrieves the dimension reduction subspace</p><p>through coherent nonparametric Bayesian kernel models. We also</p><p>discuss and provide a new rationalization of kernel regression based</p><p>on nonparametric Bayesian models allowing for direct and formal</p><p>inference on the uncertain regression functions. Our proposed models</p><p>apply for high dimensional cases where the number of variables far</p><p>exceed the sample size, and hold for both the classical setting of</p><p>Euclidean subspaces and the Riemannian setting where the marginal</p><p>distribution is concentrated on a manifold. Our Bayesian perspective</p><p>adds appropriate probabilistic and statistical frameworks that allow</p><p>for rich inference such as uncertainty estimation which is important</p><p>for measuring the estimates. Formal probabilistic models with</p><p>likelihoods and priors are given and efficient posterior sampling</p><p>can be obtained by Markov chain Monte Carlo methodologies,</p><p>particularly Gibbs sampling schemes. For the supervised dimension</p><p>reduction as the posterior draws are linear subspaces which are</p><p>points on a Grassmann manifold, we do the posterior inference with</p><p>respect to geodesics on the Grassmannian. The utility of our</p><p>approaches is illustrated on simulated and real examples.</p> / Dissertation

<p>This thesis presents spatially and temporally dependent hierarchical Bayesian models for the analysis of social networks and associated textual data. Social network analysis has received significant recent attention and has been applied to fields as varied as analysis of Supreme Court votes, Congressional roll call data, and inferring links between authors of scientific papers. In many traditional social network analysis models, temporal and spatial dependencies are not considered due to computational difficulties, even though significant such dependencies often play a significant role in the underlying generative process of the observed social network data.</p><p>Thus motivated, this thesis presents four new models that consider spatial and/or temporal dependencies and (when available) the associated text. The first is a time-dependent (dynamic) relational topic model that models nodes by their relevant documents and uses probit regression construction to map topic overlap between nodes to a link. The second is a factor model with dynamic random effects that is used to analyze the voting patterns of the United States Supreme Court. hTe last two models present the primary contribution of this thesis two spatially and temporally dependent models that jointly analyze legislative roll call data and the their associated legislative text and introduce a new paradigm for social network factor analysis: being able to predict new columns (or rows) of matrices from the text. The first uses a nonparametric joint clustering approach to link the factor and topic models while the second uses a text regression construction. Finally, two other models on analysis of and tracking in video are also presented and discussed.</p> / Dissertation

The protein structure prediction problem consists of determining a protein’s three-dimensional structure from the underlying sequence of amino acids. A standard approach for predicting such structures is to conduct a stochastic search of conformation space in an attempt to find a conformation that optimizes a scoring function. For one subclass of prediction protocols, called template-based modeling, a new protein is suspected to be structurally similar to other proteins with known structure. The solved related proteins may be used to guide the search of protein structure space. There are many potential applications for statistics in this area, ranging from the development of structure scores to improving search algorithms. This dissertation focuses on strategies for improving structure predictions by incorporating information about closely related “template” protein structures into searches of protein conformation space. This is accomplished by generating density estimates on conformation space via various simplifications of structure models. By concentrating a search for good structure conformations in areas that are inhabited by similar proteins, we improve the efficiency of our search and increase the chances of finding a low-energy structure. In the course of addressing this structural biology problem, we present a number of advances to the field of Bayesian nonparametric density estimation. We first develop a method for density estimation with bivariate angular data that has applications to characterizing protein backbone conformation space. We then extend this model to account for multiple angle pairs, thereby addressing the problem of modeling protein regions instead of single sequence positions. In the course of this analysis we incorporate an informative prior into our nonparametric density estimate and find that this significantly improves performance for protein loop prediction. The final piece of our structure prediction strategy is to connect side-chain locations to our torsion angle representation of the protein backbone. We accomplish this by using a Bayesian nonparametric model for dependence that can link together two or more multivariate marginals distributions. In addition to its application for our angular-linear data distribution, this dependence model can serve as an alternative to nonparametric copula methods.