Probabilistic Structured Models for Plant Trait Analysis

Abstract

Many fields in modern science and engineering such as ecology, computational biology, astronomy, signal processing, climate science, brain imaging, natural language processing, and many more involve collecting data sets in which the dimensionality of the data p exceeds the sample size n. Since it is usually impossible to obtain consistent procedures unless p < n, a line of recent work has studied models with various types of low-dimensional structure, including sparse vectors, sparse structured graphical models, low-rank matrices, and combinations thereof. In such settings, a general approach to estimation is to solve a regularized optimization problem, which combines a loss function measuring how well the model fits the data with some regularization function that encourages the assumed structure. Of particular interest are structure learning of graphical models in high dimensional setting. The majority of statistical analysis of graphical model estimations assume that all the data are fully observed and the data points are sampled from the same distribution and provide the sample complexity and convergence rate by considering only one graphical structure for all the observations. In this thesis, we extend the above results to estimate the structure of graphical models where the data is partially observed or the data is sampled from multiple distributions. First, we consider the problem of estimating change in the dependency structure of two p-dimensional models, based on samples drawn from two graphical models. The change is assumed to be structured, e.g., sparse, block sparse, node-perturbed sparse, etc., such that it can be characterized by a suitable (atomic) norm. We present and analyze a norm-regularized estimator for directly estimating the change in structure, without having to estimate the structures of the individual graphical models. Next, we consider the problem of estimating sparse structure of Gaussian copula distributions (corresponding to non-paranormal distributions) using samples with missing values. We prove that our proposed estimators consistently estimate the non-paranormal correlation matrix where the convergence rate depends on the probability of missing values. In the second part of thesis, we consider matrix completion problem. Low-rank matrix completion methods have been successful in a variety of settings such as recommendation systems. However, most of the existing matrix completion methods only provide a point estimate of missing entries, and do not characterize uncertainties of the predictions. First, we illustrate that the the posterior distribution in latent factor models, such as probabilistic matrix factorization, when marginalized over one latent factor has the Matrix Generalized Inverse Gaussian (MGIG) distribution. We show that the MGIG is unimodal, and the mode can be obtained by solving an Algebraic Riccati Equation equation. The characterization leads to a novel Collapsed Monte Carlo inference algorithm for such latent factor models. Next, we propose a Bayesian hierarchical probabilistic matrix factorization (BHPMF) model to 1) incorporate hierarchical side information, and 2) provide uncertainty quantified predictions. The former yields significant performance improvements in the problem of plant trait prediction, a key problem in ecology, by leveraging the taxonomic hierarchy in the plant kingdom. The latter is helpful in identifying predictions of low confidence which can in turn be used to guide field work for data collection efforts. Finally, we consider applications of probabilistic structured models to plant trait analysis. We apply BHPMF model to fill the gaps in TRY database. The BHPMF model is the-state-of-the-art model for plant trait prediction and is getting increasing visibility and usage in the plant trait analysis. We have submitted a R package for BHPMF to CRAN. Next, we apply the Gaussian graphical model structure estimators to obtain the trait-trait interactions. We study the trait-trait interactions structure at different climate zones and among different plant growth forms and uncover the dependence of traits on climate and on vegetation.