<p>Graphs are the natural framework of many of today’s highest impact computing applications: from online social networking, to Web search, to product recommendations, to chemistry, to bioinformatics, to knowledge bases, to mobile ad-hoc networking. To develop successful applications in these domains, we often need representation learning methods ---models mapping nodes, edges, subgraphs or entire graphs to some meaningful vector space. Such models are studied in the machine learning subfield of graph representation learning (GRL). Previous GRL research has focused on learning node or entire graph representations through associational tasks. In this work I study higher-order (k>1-node) representations of graphs in the context of both associational and counterfactual tasks.<br> </p>

This thesis studies realized volatility (RV), implied volatility (IV) and their applications in stochastic volatility models. The first essay uses both daytime and overnight high-frequency price data for equity index futures to estimate the RV of the S\&P500 and NASDAQ 100 indexes. Empirical results reveal strong inter-correlation between the regular-trading-time and after-hour RVs, as well as a significant predictive power of overnight RV on daytime RV and vice versa. We propose a new day-night realized stochastic volatility (DN-SV-RV) model, where the daytime and overnight returns are jointly modeled with their RVs, and their latent volatilities are correlated. The newly proposed DN-SV-RV model has the best out-of-sample return distribution forecasts among the models considered. The second essay extends the realized stochastic volatility model by jointly estimating return, RV and IV. We examine how RV and IV enhance the estimation of the latent volatility process for both the S\&P500 index and individual stocks. The third essay re-examines asymmetric stochastic volatility (ASV) models with different return-volatility correlation structures given RV and IV. We show by simulation that estimating the ASV models with return series alone may infer erroneous estimations of the correlation coefficients. The incorporation of volatility measures helps identify the true return-volatility correlation within the ASV framework. Empirical evidence on global equity market indices verifies that ASV models with additional volatility measures not only obtain significantly different estimations of the correlations compared to the benchmark ASV models, but also improve out-of-sample return forecasts. / Thesis / Doctor of Philosophy (PhD)

The rapid development of biotechnology has enabled researchers to collect high-throughput data for studying various biological processes at the genomic level, transcriptomic level, and proteomic level. Due to the large noise in the data and the high complexity of diseases (such as cancer), it is a challenging task for researchers to extract biologically meaningful information that can help reveal the underlying molecular mechanisms. The challenges call for more efforts in developing efficient and effective computational methods to analyze the data at different levels so as to understand the biological systems in different aspects. In this dissertation research, we have developed novel Bayesian approaches to infer alternative splicing mechanisms in biological systems using RNA sequencing data. Specifically, we focus on two research topics in this dissertation: isoform identification and phenotype-specific transcript assembly. For isoform identification, we develop a computational approach, SparseIso, to jointly model the existence and abundance of isoforms in a Bayesian framework. A spike-and-slab prior is incorporated into the model to enforce the sparsity of expressed isoforms. A Gibbs sampler is developed to sample the existence and abundance of isoforms iteratively. For transcript assembly, we develop a Bayesian approach, IntAPT, to assemble phenotype-specific transcripts from multiple RNA sequencing profiles. A two-layer Bayesian framework is used to model the existence of phenotype-specific transcripts and the transcript abundance in individual samples. Based on the hierarchical Bayesian model, a Gibbs sampling algorithm is developed to estimate the joint posterior distribution for phenotype-specific transcript assembly. The performances of our proposed methods are evaluated with simulation data, compared with existing methods and benchmarked with real cell line data. We then apply our methods on breast cancer data to identify biologically meaningful splicing mechanisms associated with breast cancer. For the further work, we will extend our methods for de novo transcript assembly to identify novel isoforms in biological systems; we will incorporate isoform-specific networks into our methods to better understand splicing mechanisms in biological systems. / Ph. D.

Computational systems biology is an inter-disciplinary field that aims to develop computational approaches for a system-level understanding of biological systems. Advances in high-throughput biotechnology offer broad scope and high resolution in multiple disciplines. However, it is still a major challenge to extract biologically meaningful information from the overwhelming amount of data generated from biological systems. Effective computational approaches are of pressing need to reveal the functional components. Thus, in this dissertation work, we aim to develop computational approaches for differential analysis of RNA-seq and methylation data to detect aberrant events associated with cancers. We develop a novel Bayesian approach, BayesIso, to identify differentially expressed isoforms from RNA-seq data. BayesIso features a joint model of the variability of RNA-seq data and the differential state of isoforms. BayesIso can not only account for the variability of RNA-seq data but also combines the differential states of isoforms as hidden variables for differential analysis. The differential states of isoforms are estimated jointly with other model parameters through a sampling process, providing an improved performance in detecting isoforms of less differentially expressed. We propose to develop a novel probabilistic approach, DM-BLD, in a Bayesian framework to identify differentially methylated genes. The DM-BLD approach features a hierarchical model, built upon Markov random field models, to capture both the local dependency of measured loci and the dependency of methylation change. A Gibbs sampling procedure is designed to estimate the posterior distribution of the methylation change of CpG sites. Then, the differential methylation score of a gene is calculated from the estimated methylation changes of the involved CpG sites and the significance of genes is assessed by permutation-based statistical tests. We have demonstrated the advantage of the proposed Bayesian approaches over conventional methods for differential analysis of RNA-seq data and methylation data. The joint estimation of the posterior distributions of the variables and model parameters using sampling procedure has demonstrated the advantage in detecting isoforms or methylated genes of less differential. The applications to breast cancer data shed light on understanding the molecular mechanisms underlying breast cancer recurrence, aiming to identify new molecular targets for breast cancer treatment. / Ph. D.

Mathematical models of viral infections have been informing virology research for years. Estimating parameter values for these models can lead to understanding of biological values. This has been successful in HIV modeling for the estimation of values such as the lifetime of infected CD8 T-Cells. However, estimating these values is notoriously difficult, especially for highly complex models. We use Bayesian inference and Monte Carlo Markov Chain methods to estimate the underlying densities of the parameters (assumed to be continuous random variables) for three models of influenza. We discuss the advantages and limitations of parameter estimation using these methods. The data and influenza models used for this project are from the lab of Dr. Amber Smith in Memphis, Tennessee. / Master of Science

Liquid chromatography coupled with mass spectrometry (LC-MS) has been widely used in various omic studies for biomarker discovery. Appropriate LC-MS data preprocessing steps are needed to detect true differences between biological groups. Retention time alignment is one of the most important yet challenging preprocessing steps, in order to ensure that ion intensity measurements among multiple LC-MS runs are comparable. In this dissertation, we propose a Bayesian alignment model (BAM) for analysis of LC-MS data. BAM uses Markov chain Monte Carlo (MCMC) methods to draw inference on the model parameters and provides estimates of the retention time variability along with uncertainty measures, enabling a natural framework to integrate information of various sources. From methodology development to practical application, we investigate the alignment problem through three research topics: 1) development of single-profile Bayesian alignment model, 2) development of multi-profile Bayesian alignment model, and 3) application to biomarker discovery research. Chapter 2 introduces the profile-based Bayesian alignment using a single chromatogram, e.g., base peak chromatogram from each LC-MS run. The single-profile alignment model improves on existing MCMC-based alignment methods through 1) the implementation of an efficient MCMC sampler using a block Metropolis-Hastings algorithm, and 2) an adaptive mechanism for knot specification using stochastic search variable selection (SSVS). Chapter 3 extends the model to integrate complementary information that better captures the variability in chromatographic separation. We use Gaussian process regression on the internal standards to derive a prior distribution for the mapping functions. In addition, a clustering approach is proposed to identify multiple representative chromatograms for each LC-MS run. With the Gaussian process prior, these chromatograms are simultaneously considered in the profile-based alignment, which greatly improves the model estimation and facilitates the subsequent peak matching process. Chapter 4 demonstrates the applicability of the proposed Bayesian alignment model to biomarker discovery research. We integrate the proposed Bayesian alignment model into a rigorous preprocessing pipeline for LC-MS data analysis. Through the developed analysis pipeline, candidate biomarkers for hepatocellular carcinoma (HCC) are identified and confirmed on a complementary platform. / Ph. D.

We consider the statistical inverse problem of estimating a background flow field (e.g., of air or water) from the partial and noisy observation of a passive scalar (e.g., the concentration of a pollutant). Here the unknown is a vector field that is specified by large or infinite number of degrees of freedom. We show that the inverse problem is ill-posed, i.e., there may be many or no background flows that match a given set of observations. We therefore adopt a Bayesian approach, incorporating prior knowledge of background flows and models of the observation error to develop probabilistic estimates of the fluid flow. In doing so, we leverage frameworks developed in recent years for infinite-dimensional Bayesian inference. We provide conditions under which the inference is consistent, i.e., the posterior measure converges to a Dirac measure on the true background flow as the number of observations of the solute concentration grows large. We also define several computationally-efficient algorithms adapted to the problem. One is an adjoint method for computation of the gradient of the log likelihood, a key ingredient in many numerical methods. A second is a particle method that allows direct computation of point observations of the solute concentration, leveraging the structure of the inverse problem to avoid approximation of the full infinite-dimensional scalar field. Finally, we identify two interesting example problems with very different posterior structures, which we use to conduct a large-scale benchmark of the convergence of several Markov Chain Monte Carlo methods that have been developed in recent years for infinite-dimensional settings. / Ph. D.

Les tâches de vision artificielle telles que la reconnaissance d’objets demeurent irrésolues à ce jour. Les algorithmes d’apprentissage tels que les Réseaux de Neurones Artificiels (RNA), représentent une approche prometteuse permettant d’apprendre des caractéristiques utiles pour ces tâches. Ce processus d’optimisation est néanmoins difficile. Les réseaux profonds à base de Machine de Boltzmann Restreintes (RBM) ont récemment été proposés afin de guider l’extraction de représentations intermédiaires, grâce à un algorithme d’apprentissage non-supervisé. Ce mémoire présente, par l’entremise de trois articles, des contributions à ce domaine de recherche. Le premier article traite de la RBM convolutionelle. L’usage de champs réceptifs locaux ainsi que le regroupement d’unités cachées en couches partageant les même paramètres, réduit considérablement le nombre de paramètres à apprendre et engendre des détecteurs de caractéristiques locaux et équivariant aux translations. Ceci mène à des modèles ayant une meilleure vraisemblance, comparativement aux RBMs entraînées sur des segments d’images. Le deuxième article est motivé par des découvertes récentes en neurosciences. Il analyse l’impact d’unités quadratiques sur des tâches de classification visuelles, ainsi que celui d’une nouvelle fonction d’activation. Nous observons que les RNAs à base d’unités quadratiques utilisant la fonction softsign, donnent de meilleures performances de généralisation. Le dernière article quand à lui, offre une vision critique des algorithmes populaires d’entraînement de RBMs. Nous montrons que l’algorithme de Divergence Contrastive (CD) et la CD Persistente ne sont pas robustes : tous deux nécessitent une surface d’énergie relativement plate afin que leur chaîne négative puisse mixer. La PCD à "poids rapides" contourne ce problème en perturbant légèrement le modèle, cependant, ceci génère des échantillons bruités. L’usage de chaînes tempérées dans la phase négative est une façon robuste d’adresser ces problèmes et mène à de meilleurs modèles génératifs. / High-level vision tasks such as generic object recognition remain out of reach for modern Artificial Intelligence systems. A promising approach involves learning algorithms, such as the Arficial Neural Network (ANN), which automatically learn to extract useful features for the task at hand. For ANNs, this represents a difficult optimization problem however. Deep Belief Networks have thus been proposed as a way to guide the discovery of intermediate representations, through a greedy unsupervised training of stacked Restricted Boltzmann Machines (RBM). The articles presented here-in represent contributions to this field of research. The first article introduces the convolutional RBM. By mimicking local receptive fields and tying the parameters of hidden units within the same feature map, we considerably reduce the number of parameters to learn and enforce local, shift-equivariant feature detectors. This translates to better likelihood scores, compared to RBMs trained on small image patches. In the second article, recent discoveries in neuroscience motivate an investigation into the impact of higher-order units on visual classification, along with the evaluation of a novel activation function. We show that ANNs with quadratic units using the softsign activation function offer better generalization error across several tasks. Finally, the third article gives a critical look at recently proposed RBM training algorithms. We show that Contrastive Divergence (CD) and Persistent CD are brittle in that they require the energy landscape to be smooth in order for their negative chain to mix well. PCD with fast-weights addresses the issue by performing small model perturbations, but may result in spurious samples. We propose using simulated tempering to draw negative samples. This leads to better generative models and increased robustness to various hyperparameters.

L'un des modèles d'apprentissage non-supervisé générant le plus de recherche active est la machine de Boltzmann --- en particulier la machine de Boltzmann restreinte, ou RBM. Un aspect important de l'entraînement ainsi que l'exploitation d'un tel modèle est la prise d'échantillons. Deux développements récents, la divergence contrastive persistante rapide (FPCD) et le herding, visent à améliorer cet aspect, se concentrant principalement sur le processus d'apprentissage en tant que tel. Notamment, le herding renonce à obtenir un estimé précis des paramètres de la RBM, définissant plutôt une distribution par un système dynamique guidé par les exemples d'entraînement. Nous généralisons ces idées afin d'obtenir des algorithmes permettant d'exploiter la distribution de probabilités définie par une RBM pré-entraînée, par tirage d'échantillons qui en sont représentatifs, et ce sans que l'ensemble d'entraînement ne soit nécessaire. Nous présentons trois méthodes: la pénalisation d'échantillon (basée sur une intuition théorique) ainsi que la FPCD et le herding utilisant des statistiques constantes pour la phase positive. Ces méthodes définissent des systèmes dynamiques produisant des échantillons ayant les statistiques voulues et nous les évaluons à l'aide d'une méthode d'estimation de densité non-paramétrique. Nous montrons que ces méthodes mixent substantiellement mieux que la méthode conventionnelle, l'échantillonnage de Gibbs. / One of the most active topics of research in unsupervised learning is the Boltzmann machine --- particularly the Restricted Boltzmann Machine or RBM. In order to train, evaluate or exploit such models, one has to draw samples from it. Two recent algorithms, Fast Persistent Contrastive Divergence (FPCD) and Herding aim to improve sampling during training. In particular, herding gives up on obtaining a point estimate of the RBM's parameters, rather defining the model's distribution with a dynamical system guided by training samples. We generalize these ideas in order to obtain algorithms capable of exploiting the probability distribution defined by a pre-trained RBM, by sampling from it, without needing to make use of the training set. We present three methods: Sample Penalization, based on a theoretical argument as well as FPCD and Herding using constant statistics for their positive phases. These methods define dynamical systems producing samples with the right statistics and we evaluate them using non-parametric density estimation. We show that these methods mix substantially better than Gibbs sampling, which is the conventional sampling method used for RBMs.

Dans cette thèse, on propose des techniques d'inférence bayésienne dans les modèles où la vraisemblance possède une composante latente. La vraisemblance d'un jeu de données observé est l'intégrale de la vraisemblance dite complète sur l'espace de la variable latente. On s'intéresse aux cas où l'espace de la variable latente est de très grande dimension et comportes des directions de différentes natures (discrètes et continues), ce qui rend cette intégrale incalculable. Le champs d'application privilégié de cette thèse est l'inférence dans les modèles de génétique des populations. Pour mener leurs études, les généticiens des populations se basent sur l'information génétique extraite des populations du présent et représente la variable observée. L'information incluant l'histoire spatiale et temporelle de l'espèce considérée est inaccessible en général et représente la composante latente. Notre première contribution dans cette thèse suppose que la vraisemblance peut être évaluée via une approximation numériquement coûteuse. Le schéma d'échantillonnage préférentiel adaptatif et multiple (AMIS pour Adaptive Multiple Importance Sampling) de Cornuet et al. [2012] nécessite peu d'appels au calcul de la vraisemblance et recycle ces évaluations. Cet algorithme approche la loi a posteriori par un système de particules pondérées. Cette technique est conçue pour pouvoir recycler les simulations obtenues par le processus itératif (la construction séquentielle d'une suite de lois d'importance). Dans les nombreux tests numériques effectués sur des modèles de génétique des populations, l'algorithme AMIS a montré des performances numériques très prometteuses en terme de stabilité. Ces propriétés numériques sont particulièrement adéquates pour notre contexte. Toutefois, la question de la convergence des estimateurs obtenus parcette technique reste largement ouverte. Dans cette thèse, nous montrons des résultats de convergence d'une version légèrement modifiée de cet algorithme. Sur des simulations, nous montrons que ses qualités numériques sont identiques à celles du schéma original. Dans la deuxième contribution de cette thèse, on renonce à l'approximation de la vraisemblance et onsupposera seulement que la simulation suivant le modèle (suivant la vraisemblance) est possible. Notre apport est un algorithme ABC séquentiel (Approximate Bayesian Computation). Sur les modèles de la génétique des populations, cette méthode peut se révéler lente lorsqu'on vise uneapproximation précise de la loi a posteriori. L'algorithme que nous proposons est une amélioration de l'algorithme ABC-SMC de DelMoral et al. [2012] que nous optimisons en nombre d'appels aux simulations suivant la vraisemblance, et que nous munissons d'un mécanisme de choix de niveauxd'acceptations auto-calibré. Nous implémentons notre algorithme pour inférer les paramètres d'un scénario évolutif réel et complexe de génétique des populations. Nous montrons que pour la même qualité d'approximation, notre algorithme nécessite deux fois moins de simulations par rapport à laméthode ABC avec acceptation couramment utilisée. / This thesis consists of two parts which can be read independently.The first part is about the Adaptive Multiple Importance Sampling (AMIS) algorithm presented in Cornuet et al.(2012) provides a significant improvement in stability and Effective Sample Size due to the introduction of the recycling procedure. These numerical properties are particularly adapted to the Bayesian paradigm in population genetics where the modelization involves a large number of parameters. However, the consistency of the AMIS estimator remains largely open. In this work, we provide a novel Adaptive Multiple Importance Sampling scheme corresponding to a slight modification of Cornuet et al. (2012) proposition that preserves the above-mentioned improvements. Finally, using limit theorems on triangular arrays of conditionally independant random variables, we give a consistensy result for the final particle system returned by our new scheme.The second part of this thesis lies in ABC paradigm. Approximate Bayesian Computation has been successfully used in population genetics models to bypass the calculation of the likelihood. These algorithms provide an accurate estimator by comparing the observed dataset to a sample of datasets simulated from the model. Although parallelization is easily achieved, computation times for assuring a suitable approximation quality of the posterior distribution are still long. To alleviate this issue, we propose a sequential algorithm adapted fromDel Moral et al. (2012) which runs twice as fast as traditional ABC algorithms. Itsparameters are calibrated to minimize the number of simulations from the model.