Beyond Domain Boundaries: Enhancing The Generalizability of Event Sequence Analytics

Abstract

Event sequence data consists of chronologically arranged records of timestamped events connected to entities, systems, or processes. It is present across a broad range of domains such as healthcare, industrial operations, and digital marketing. Practitioners can identify behavioral trends, forecast outcomes, and aid in decision-making by analyzing such data. However, the current landscape of event sequence visual analytics is highly fragmented. Existing research heavily prioritizes the development of tightly integrated, domain-specific systems that are evaluated exclusively through domain-specific case studies. As a result, the field is rich in domain-specific solutions but poor in generalizable principles.

This dissertation addresses this fragmentation by establishing methodologies to enhance the generalizability of event sequence visual analytics. We define generalizability as the ability to decouple analytical methods and visual representations from their specific contextual dependencies, allowing them to be broadly applied. To transition the field from isolated solutions to a unified research space, this work contributes along three fundamental dimensions: theoretical, computational, and empirical.

First, to address the lack of a shared analytical vocabulary, we present a theoretical contribution: a domain-agnostic, multi level task framework. Derived from an extensive analysis of existing visualization systems, this framework abstracts event sequence analysis into four hierarchical levels: objectives, intents, strategies, and techniques. By decoupling the fundamental analytical goals from the specific data they operate on, this framework establishes a universal language for the field. The framework enables researchers and practitioners to recognize analytically equivalent problems across domains, evaluate tools on a standardized basis, and transfer knowledge between disparate application areas.

Second, we introduce a computational contribution to handle real-world non-stationarity sequence generation conditions. We propose a domain-agnostic causal framework that extends classical treatment effect estimation to the temporal point process setting, formalizing the average treatment effect conditioned on out-of-domain intervention states and providing a propensity score-based estimator that accounts for confounding by observed event history. Because the modelling components: causes, outcomes, and interventions are formulated as abstract definitions instantiated with domain-specific meaning at application time, the extraction of causal relations is robust and universally applicable to event sequence data from any domain.

Finally, this dissertation provides two empirical contributions to address evaluation gaps limiting the generalizability of event sequence visualization research. We first conduct an insight-based crowdsourcing experiment to comparatively evaluate visual summarization techniques for human understanding and propose a reproducible evaluation framework. This study revealed that visual pattern complexity is a primary factor influencing comprehension. However, it also highlighted a critical methodological challenge: iteratively testing complex visual structures using crowdsourced human studies is inherently expensive, difficult to scale, and increasingly susceptible to unreliable results. To explore a scalable and reliable alternative, we introduce ProcVQA, a benchmark designed to empirically evaluate the baseline visual comprehension of Vision-Language Models (VLMs). Building directly on the findings from our human study, ProcVQA specifically evaluates how well VLMs can comprehend process visualizations across varying levels of structural complexity, providing the foundation required to utilize AI as an automated evaluation tool.

Together, these contributions establish a unified foundation for generalizable event sequence visual analytics: a shared language for describing analysis tasks, a principled model for understanding causal dynamics under real-world conditions, and standardized methods for evaluating visual representations.