ISBO
A large-scale Bayesian optimization framework for spatio-temporal point process data. It combines Log-Gaussian Cox Process modeling with INLA-SPDE sparse inference and employs a time-varying UCB acquisition function for sequential optimization.
Timeline (last 90 days)
ISBO (INLA-SPDE Spatio-Temporal Bayesian Optimization) is a scalable Bayesian optimization framework for spatio-temporal point process data that combines the Log-Gaussian Cox Process (LGCP) with the INLA-SPDE inference method.
ISBO adopts the INLA-SPDE method, using a Matérn field on a mesh to approximate a dense Gaussian process as a sparse Gaussian Markov random field (GMRF), thereby supporting fast posterior inference.
ISBO uses a time-varying upper confidence bound (UCB) acquisition function and adds a masking mechanism to avoid repeatedly visiting already explored regions.
ISBO introduces a complexity-penalizing prior to regularize the early rounds of optimization, suppressing overconfidence or violent fluctuation in the posterior of the latent intensity field when early samples are scarce.
Experiments on synthetic and real-world spatio-temporal datasets show that ISBO can stably locate high-intensity regions and latent intensity peaks while completing optimization with fewer evaluations.
Compared with baseline methods based on RKHS (Reproducing Kernel Hilbert Space), ISBO achieves a significant speedup.
ISBO claims to be the first scalable spatiotemporal Bayesian optimization framework.
ISBO is applicable to spatiotemporal data scenarios that appear in the form of event points, such as epidemic spread, crime hotspots, traffic incidents, and ecological observations.
All Facts (8)
ISBO is applicable to spatiotemporal data scenarios that appear in the form of event points, such as epidemic spread, crime hotspots, traffic incidents, and ecological observations.
50%UnverifiedISBO adopts the INLA-SPDE method, using a Matérn field on a mesh to approximate a dense Gaussian process as a sparse Gaussian Markov random field (GMRF), thereby supporting fast posterior inference.
50%UnverifiedISBO uses a time-varying upper confidence bound (UCB) acquisition function and adds a masking mechanism to avoid repeatedly visiting already explored regions.
50%UnverifiedISBO introduces a complexity-penalizing prior to regularize the early rounds of optimization, suppressing overconfidence or violent fluctuation in the posterior of the latent intensity field when early samples are scarce.
50%UnverifiedExperiments on synthetic and real-world spatio-temporal datasets show that ISBO can stably locate high-intensity regions and latent intensity peaks while completing optimization with fewer evaluations.
50%UnverifiedCompared with baseline methods based on RKHS (Reproducing Kernel Hilbert Space), ISBO achieves a significant speedup.
50%UnverifiedISBO claims to be the first scalable spatiotemporal Bayesian optimization framework.
50%UnverifiedISBO (INLA-SPDE Spatio-Temporal Bayesian Optimization) is a scalable Bayesian optimization framework for spatio-temporal point process data that combines the Log-Gaussian Cox Process (LGCP) with the INLA-SPDE inference method.
50%