Title: Nonparametric clustering for spatio-temporal data
Authors: Ashwini Venkatasubramaniam - Alan Turing Institute (United Kingdom) [presenting]
Konstantinos Ampountolas - University of Glasgow (United Kingdom)
Ludger Evers - University of Glasgow (United Kingdom)
Abstract: A non-parametric clustering approach for spatio-temporal dataset is proposed which seeks to identify spatially contiguous clusters and retain the underlying temporal patterns for associated clusters. This flexible Bayesian method utilises a modified distance dependent Chinese restaurant process (ddCRP), referred to as the netCRP, to model network connectivity and spatial dependencies. The netCRP seeks to incorporate neighbourhood relationships for each vertex in the graph and the graph network in this context is assumed to be composed of vertices that have a limited number of adjacent vertices. The non-sequential ddCRP is modified to also allow for the ability to control the number of self-links and redundant links between vertices; individual clusters in the network are formed by the formation of cycles. In order to fully account for within-cluster spatial and temporal correlation, the model defines a spatio-temporal precision matrix using a type of conditional auto-regressive (CAR) model and first order auto-regressive (AR-1) model. The model utilises a Metropolis within Gibbs sampler to fully explore all possible cluster configurations in the network and infer the relevant parameters. We illustrate this developed clustering method using applications to grid-style and map-based graph networks.