UC Davis

Network data is pervasive in daily life. Various associations among entities found in real-world applications can be represented as multivariate networks, temporal networks, or heterogeneous networks. However, as data complexity increases, analyzing networks becomes a challenging task. To draw actionable insights, it is important to have the ability to uncover and reason network characteristics associated with behaviors/events of interest. In recent years, machine learning (ML) has proven effective in network feature extraction, while visual analytics (VA) has shown value in dissecting and reasoning complex data. A promising approach to complex network analysis is to combine these two techniques.This dissertation aims to introduce new ML-assisted visual analytics designs for understanding complex networks. Firstly, to address the challenges of reasoning multivariate networks, I develop a visual analytics system that supports users in selecting network variables of interest and flexibly constructing a composite variable. This composite variable is generated through a series of learning-based optimizations and visually presented to convey the combined associations between a group of variables and another variable of interest.Secondly, driven by the desire to improve patient care quality, I conduct longitudinal studies on healthcare professionals (HCPs) collaboration networks constructed from patient records. A visual analytics system is designed to investigate the effectiveness and efficiency of each such temporal teamwork network pertinent to the respective cancer patient. This work aims to associate the structural features and dynamics of the HCPs' communication network with patient survival outcomes using network measures. Users can interactively define the analysis timeframe and visually explore how information is transmitted among HCPs and determine whether each piece of patient information has been effectively and efficiently communicated throughout the patient care process.Third, beyond detecting associations between network measures (e.g., edge density) and patient survival outcomes, identifying the most relevant collaborative patterns can offer a more intuitive understanding. Graph Neural Networks (GNNs) may be used to extract structural features from complex networks but their complicated data transformation poses challenges in explaining their behavior. To address this shortcoming, I devise a visual analytics system to evaluate and explain the significance of each elementary substructure (i.e., graphlet) to the outputs of the GNN in graph-level classification tasks. In this system, by investigating the relationships between GNN classifications and the composition of elementary substructures, hypothesized explanations can be formulated and evaluated using metrics derived from both factual and counterfactual reasoning processes.Finally, building on these prior efforts toward interpretable machine learning for patient risk assessment and intervention planning, I develop an interactive visual analytics framework for predictive analysis and what-if simulation. The framework supports the full analytical pipeline, from data curation and mitigation of foreseeable confounding factors to predicting cancer patient morbidity risk using GNNs. It also provides methods for explaining model predictions by identifying critical network structures and key HCP participation patterns that influence decision-making. Based on these explanations, users can simulate targeted interventions and examine projected changes in patient risk. Through this framework, users can explore how collaboration dynamics relate to estimated risk, generate intervention hypotheses, and evaluate them through interactive what-if analyses.In conclusion, this dissertation advances complex network analysis by integrating machine learning and visual analytics to bridge pattern discovery, interpretation, and decision-making. Through a series of contributions spanning multivariate reasoning, temporal collaboration analysis, interpretable graph learning, and predictive what-if simulation, it demonstrates how complex network data can be transformed into actionable insights. By emphasizing transparency and human-centered reasoning, these efforts establish a foundation for trustworthy, data-driven analysis and decision support across complex networked systems.