Degree

Doctor of Philosophy (PhD)

Department

Computer Science

Document Type

Dissertation

Abstract

This research presents a comprehensive series of studies aimed at advancing learning neural point processes for long event sequences, with applications spanning disaster resilience, crime forecasting, and healthcare. Structured around three interconnected studies, this work addresses core challenges in temporal point process (TPP) modeling, efficient handling of long event sequences, and improving accuracy over extended forecasting horizons by reinforcement learning.

The first study proposes the Sparse Transformer Hawkes Process (STHP) to model long asynchronous event sequences. Traditional neural network-based TPPs struggle with long event sequences due to computational inefficiencies. To address this, the STHP model combines two components: a temporal sparse self-attention mechanism that focuses on short-term dependencies and a second transformer applied to aggregated event counts for long-term dependency extraction. By integrating these components, STHP models the conditional intensity of point processes efficiently, improving prediction performance for long sequences without high computational costs.

The second study proposes a debiased imitation learning (DIL) framework for Modulated Temporal Point Processes to handle biased event sequences encountered in applications like disaster resilience, criminology, and healthcare. In real-world settings, temporal events often suffer from unknown biases due to external factors, which lead to misspecifications in TPPs when learned using conventional maximum likelihood estimation (MLE). To address these biases, the DIL framework explicitly models biased sequences through additional unknown thinning processes, mitigating the impact of biased data. By leveraging a sequence-level reward function derived from historical embeddings, the DIL framework enhances prediction robustness and accuracy.

The third study introduces the Inflence-Guided Reinforcement Learning Spatio-Temporal Point Process (IGPO) framework, a novel framework for modeling spatio-temporal event sequences. Spatio-temporal data presents unique challenges due to dependencies across both spatial and temporal dimensions, and neural network-based spatio-temporal point processes provide a sophisticated modeling framework for these data. Conventional MLE may lead to inaccurate predictions because of model misspecification and compounding errors. Alternatively, reinforcement learning, which treats event generation as actions to mimic observed event patterns, can address the training and testing discrepancy but often suffers from poor exploration efficiency.

Together, these studies contribute a cohesive set of advancements in long event sequence prediction, addressing key challenges in efficiency, biased data, model misspecification, and compounding errors.

Date

8-24-2025

Committee Chair

Sun, Mingxuan

Download

Share

COinS