Year: 2025
Type: dissertation
Abstract: For many decades, various fields-including physics, biology, and epidemiology-have routinely used ordinary differential equations (ODEs) to describe the evolution of complex systems.A central task in this domain is the inverse problem of estimating ODE parameters from observed data, enabling both deeper understanding and more accurate predictions of underlying phenomena.Traditionally, ODE-based models assume that (i) the measured variables are continuous-valued and (ii) observations are often collected at fixed, regular time intervals.In contrast, in many real-world applications, we encounter discrete events with two fundamentally different characteristics:1. Instantaneous occurrences: each event represents a distinct, momentary occurrence or transition without duration.Building on these concepts, Chapter 3 proposes an ODE-guided Poisson process model that unifies multiple event types under one ODE-based scheme.Specifically, the intensity function for each event type evolves according to an ODE, which allows connections to domain-specific mechanisms.To make the inference of ODE parameters tractable from discrete event data, we develop a method called Log-Gaussian Cox Process based Gradient Matching (LGCP-GM).By leveraging log-Gaussian Cox processes in conjunction with Gaussian process-based gradient matching, LGCP-GM estimates both parameters and latent trajectories while keeping computational requirements reasonable.This chapter also illustrates how small-interval Poisson likelihoods, GP priors, and high-dimensional parameter sampling can be combined to keep the method scalable.A more challenging setting arises when event labels are unobserved, making it impossible to link each observed event directly to a specific ODE component.Chapter 4 addresses this by introducing an ODE-guided Poisson process allocation, which simultaneously infers both latent event types and ODE parameters.Through mark information-such as a text description attached to each event-we use a probabilistic method that assigns every observed event to a suitable type.In this context, inference in the GP-based gradient matching (GPGM) framework can become difficult due to strong dependencies between latent trajectories and ODE parameters, which hinder efficient Markov chain Monte Carlo (MCMC) sampling.To alleviate these dependencies, we introduce a Random Fourier Features-based Gradient Matching (RFFGM) approach.v By approximating the Gaussian process with finite-dimensional random feature expansions, RFFGM reduces posterior correlations, enabling more efficient MCMC sampling.The resulting inference procedure, Poisson Process Allocation-integrated Gradient Matching, thus jointly samples ODE parameters, latent functions, and hidden event assignments, extending interpretability even in cases with only indirect observations about event heterogeneity.We evaluate these methods using both synthetic data and real-world case studies.Notably, our framework recovers ODE parameters and trajectories from irregular event occurrences in canonical population-dynamic models (e.g., predator-prey and competition models, and the SIR model).We further demonstrate applications to real data-such as patent applications, web search data, and social media trend diffusion-showing that the proposed approach offers clarity regarding event-generating mechanisms and remains feasible for large datasets.By focusing on interpretable, domain-driven parameter representations, our method preserves insights that purely data-driven models may not provide.In summary, this dissertation presents a unified modeling and inference strategy for discrete event data that combines the interpretability of ODE-based methods with the adaptability of temporal point processes.By directly anchoring the temporal point process framework to ODE parameters, we achieve a clearer understanding of underlying dynamics without compromising computational efficiency.Moreover, the method also naturally extends to unlabeled events by incorporating additional mark information to infer latent event types.These contributions open new opportunities for mechanism modeling in various fields where understanding phenomena in a parameter-driven manner is often of primary importance.Future improvementssuch as integrating higher-dimensional mark data, refining hyperparameter tuning, or incorporating domain-specific knowledge-are expected to further enhance the practical impact of this approach.Dr. Kenji Fukumizu, Dr. Hideitsu Hino, Dr. Daichi Mochihashi, and Dr. Futoshi Futami-for their thoughtful suggestions and valuable feedback.My thanks extend to all members of the Mochihashi Laboratory.Their constructive comments and the camaraderie we shared made this period both demanding and rewarding.Finally,
Authors: 宮澤, 脩一, ミヤザワ, シュウイチ, MIYAZAWA, Shuichi
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7137724535
