arXiv subjects

Kangrui Wang

Publications and source records attributed to Kangrui Wang.

25 records · Page 2Linked to original sources

arXiv2020

Soft Random Graphs in Probabilistic Metric Spaces & Inter-graph Distance

We present a new method for learning Soft Random Geometric Graphs (SRGGs), drawn in probabilistic metric spaces, with the connection function of the graph defined as the marginal posterior probability of an edge random variable, given the correlation between the nodes connected by that edge. In fact, this inter-node correlation matrix is itself a random variable in our learning strategy, and we learn this by identifying each node as a random variable, measurements of which comprise a column of a given multivariate dataset. We undertake inference with Metropolis with a 2-block update scheme. The SRGG is shown to be generated by a non-homogeneous Poisson point process, the intensity of which is location-dependent. Given the multivariate dataset, likelihood of the inter-column correlation matrix is attained following achievement of a closed-form marginalisation over all inter-row correlation matrices. Distance between a pair of graphical models learnt given the respective datasets, offers the absolute correlation between the given datasets; such inter-graph distance computation is our ultimate objective, and is achieved using a newly introduced metric that resembles an uncertainty-normalised Hellinger distance between posterior probabilities of the two learnt SRGGs, given the respective datasets. Two sets of empirical illustrations on real data are undertaken, and application to simulated data is included to exemplify the effect of incorporating measurement noise in the learning of a graphical model.

stat.ME↗

arXiv2019

Multi-resolution Multi-task Gaussian Processes

We consider evidence integration from potentially dependent observation processes under varying spatio-temporal sampling resolutions and noise levels. We develop a multi-resolution multi-task (MRGP) framework while allowing for both inter-task and intra-task multi-resolution and multi-fidelity. We develop shallow Gaussian Process (GP) mixtures that approximate the difficult to estimate joint likelihood with a composite one and deep GP constructions that naturally handle biases in the mean. By doing so, we generalize and outperform state of the art GP compositions and offer information-theoretic corrections and efficient variational approximations. We demonstrate the competitiveness of MRGPs on synthetic settings and on the challenging problem of hyper-local estimation of air pollution levels across London from multiple sensing modalities operating at disparate spatio-temporal resolutions.

stat.ML↗

arXiv2018

Correlation between Multivariate Datasets, from Inter-Graph Distance computed using Graphical Models Learnt With Uncertainties

We present a method for simultaneous Bayesian learning of the correlation matrix and graphical model of a multivariate dataset, along with uncertainties in each, to subsequently compute distance between the learnt graphical models of a pair of datasets, using a new metric that approximates an uncertainty-normalised Hellinger distance between the posterior probabilities of the graphical models given the respective dataset; correlation between the pair of datasets is then computed as a corresponding affinity measure. We achieve a closed-form likelihood of the between-columns correlation matrix by marginalising over the between-row matrices. This between-columns correlation is updated first, given the data, and the graph is then updated, given the partial correlation matrix that is computed given the updated correlation, allowing for learning of the 95$\%$ Highest Probability Density credible regions of the correlation matrix and graphical model of the data. Difference made to the learnt graphical model, by acknowledgement of measurement noise, is demonstrated on a small simulated dataset, while the large human disease-symptom network--with $>8,000$ nodes--is learnt using real data. Data on vino-chemical attributes of Portuguese red and white wine samples are employed to learn with-uncertainty graphical model of each dataset, and subsequently, the distance between these learnt graphical models.

stat.ME↗

arXiv2018

Deep Bayesian Supervised Learning given Hypercuboidally-shaped, Discontinuous Data, using Compound Tensor-Variate & Scalar-Variate Gaussian Processes

We undertake Bayesian learning of the high-dimensional functional relationship between a system parameter vector and an observable, that is in general tensor-valued. The ultimate aim is Bayesian inverse prediction of the system parameters, at which test data is recorded. We attempt such learning given hypercuboidally-shaped data that displays strong discontinuities, rendering learning challenging. We model the sought high-dimensional function, with a tensor-variate Gaussian Process (GP), and use three independent ways for learning covariance matrices of the resulting likelihood, which is Tensor-Normal. We demonstrate that the discontinuous data demands that implemented covariance kernels be non-stationary--achieved by modelling each kernel hyperparameter, as a function of the sample function of the invoked tensor-variate GP. Each such function can be shown to be temporally-evolving, and treated as a realisation from a distinct scalar-variate GP, with covariance described adaptively by collating information from a historical set of samples of chosen sample-size. We prove that deep-learning using 2-"layers", suffice, where the outer-layer comprises the tensor-variate GP, compounded with multiple scalar-variate GPs in the "inner-layer", and undertake inference with Metropolis-within-Gibbs. We apply our method to a cuboidally-shaped, discontinuous, real dataset, and subsequently perform forward prediction to generate data from our model, given our results--to perform model-checking.

stat.ME↗

arXiv2015

Bayesian Covariance Modelling of Large Tensor-Variate Data Sets $\&$ Inverse Non-parametric Learning of the Unknown Model Parameter Vector

We present a method for modelling the covariance structure of tensor-variate data, with the ulterior aim of learning an unknown model parameter vector using such data. We express the high-dimensional observable as a function of this sought model parameter vector, and attempt to learn such a high-dimensional function given training data, by modelling it as a realisation from a tensor-variate Gaussian Process (GP). The likelihood of the unknowns given training data, is then tensor-normal. We choose vague priors on the unknown GP parameters (mean tensor and covariance matrices) and write the posterior probability density of these unknowns given the data. We perform posterior sampling using Random-Walk Metropolis-Hastings. Thereafter we learn the aforementioned unknown model parameter vector by performing posterior sampling in two different ways, given test and training data, using MCMC, to generate 95$\%$ HPD credible region on each unknown. We make an application of this method to the learning of the location of the Sun in the Milky Way disk.

stat.AP↗

arXiv2015

Bayesian Covariance Modelling of Large Tensor-Variate Data Sets $\&$ Inverse Non-parametric Learning of the Unknown Model Parameter Vector

Tensor-valued data are being encountered increasingly more commonly, in the biological, natural as well as the social sciences. The learning of the unknown model parameter vector given such data, involves covariance modelling of such data, though this can be difficult owing to the high-dimensional nature of the data, where the numerical challenge of such modelling can only be compounded by the largeness of the available data set. Assuming such data to be modelled using a correspondingly high-dimensional Gaussian Process (${\cal GP}$), the joint density of a finite set of such data sets is then a tensor normal distribution, with density parametrised by a mean tensor $\boldsymbol{M}$ (that is of the same dimensionality as the $k$-tensor valued observable), and the $k$ covariance matrices $\boldsymbolΣ_1,...,\boldsymbolΣ_k$. When aiming to model the covariance structure of the data, we need to estimate/learn $\{\boldsymbolΣ_1,...,\boldsymbolΣ_k \}$ and $\boldsymbol{M}$, given tha data. We present a new method in which we perform such covariance modelling by first expressing the probability density of the available data sets as tensor-normal. We then invoke appropriate priors on these unknown parameters and express the posterior of the unknowns given the data. We sample from this posterior using an appropriate variant of Metropolis Hastings. Since the classical MCMC is time and resource intensive in high-dimensional state spaces, we use an efficient variant of the Metropolis-Hastings algorithm--Transformation based MCMC--employed to perform efficient sampling from a high-dimensional state space. Once we perform the covariance modelling of such a data set, we will learn the unknown model parameter vector at which a measured (or test) data set has been obtained, given the already modelled data (training data), augmented by the test data.

stat.AP↗

arXiv2015

Uncertainty in Test Score Data and Classically Defined Reliability of Tests and Test Batteries, using a New Method for Test Dichotomisation

As with all measurements, the measurement of examinee ability, in terms of scores that the examinee obtains in a test, is also error-ridden. The quantification of such error or uncertainty in the test score data--or rather the complementary test reliability--is pursued within the paradigm of Classical Test Theory in a variety of ways, with no existing method of finding reliability, isomorphic to the theoretical definition that parametrises reliability as the ratio of the true score variance and observed (i.e. error-ridden) score variance. Thus, multiple reliability coefficients for the same test have been advanced. This paper describes a much needed method of obtaining reliability of a test as per its theoretical definition, via a single administration of the test, by using a new fast method of splitting of a given test into parallel halves, achieving near-coincident empirical distributions of the two halves. The method has the desirable property of achieving splitting on the basis of difficulty of the questions (or items) that constitute the test, thus allowing for fast computation of reliability even for very large test data sets, i.e. test data obtained by a very large examinee sample. An interval estimate for the true score is offered, given an examinee score, subsequent to the determination of the test reliability. This method of finding test reliability as per the classical definition can be extended to find reliability of a set or battery of tests; a method for determination of the weights implemented in the computation of the weighted battery score is discussed. We perform empirical illustration of our method on real and simulated tests, and on a real test battery comprising two constituent tests.

stat.AP↗