IEEE Pacific Visualization Symposium : [proceedings]

Scatterplots are widely used in Explainable Artificial Intelligence (XAI) to investigate misclassifications and patterns across instances. However, a significant limitation of scatterplots is overplotting, especially when working with large datasets. Although several quality metrics have been proposed to measure the degree of overplotting, none have been demonstrated to be effective in the context of XAI. This paper aims to evaluate the effectiveness of a quality metric, called OM4AnI, in XAI scenarios. We begin by summarizing two visual patterns—cluster-based and regression-based patterns—that support three common XAI tasks: feature importance, feature dependency, and model accuracy. We also introduce how to select the parameters of OM4AnI based on these patterns. We construct two case studies to identify the effectiveness of OM4AnI using public datasets: Census Income dataset and MNIST dataset. OM4AnI is applied to both scenarios under various visual conditions (e.g., marker size and rendering order) to assess its effectiveness. The results demonstrate that OM4AnI serves as an effective quality metric for these two common XAI scenarios, paving the way for adapting other quality metrics to be scalable within XAI contexts.

Large Language Models (LLMs) are becoming a key part of interactive visual analytics systems, helping users engage in natural language interaction and explore data more effectively. However, this kind of support can sometimes unintentionally reinforce confirmation bias by consistently aligning responses with what the user has already expected and focused on. In this work, we introduce an event-driven workflow to raise awareness of confirmation bias, leveraging naturally occurring interaction traces (e.g., navigation patterns, content choices, time spent on items, and query formulation) to detect signs of confirmation bias during exploration. We apply our proposed method in the realm of Consumer Health Information Systems (CHIS), where mitigating biased information seeking and processing is particularly critical. Our work lays the groundwork for future research on cognitive bias-aware LLM agents, extending beyond confirmation bias to other forms of cognitive biases in visual analytics systems.

Multimodal packet streams—video synchronized with telemetry, pose estimates, and runtime events—increasingly support real-time monitoring in robotics, digital twins, and sensor networks. Frequency-domain diagnostics are attractive in this setting because they can localize where a rendered view deviates from the observation, yet they are often validated under simplified previous-frame proxies whose conclusions may not transfer to real rendering pipelines. We present Coherent Frequency Packet Splatting for Dynamics (CFPS-DYN), a visual analytics methodology that computes an Explainable Frequency Heatmap via a tiled, phase-sensitive spectral residual between the incoming observation and the current prediction. We evaluate five diagnostic representations—pixel residual, FFT amplitude, phase coherence, and two packet-based variants—under both a previous-frame proxy and real 3D Gaussian Splatting (3DGS) renders on TUM RGB-D sequences. Under the proxy, the pixel baseline exhibits an inflated correlation with temporal instability (r=0.920), but this oracle advantage collapses under real 3DGS renders (r=0.041) when temporal information is no longer leaked into the residual. We empirically observe a decoupling between temporal instability and reconstruction fidelity under real 3DGS renders, with stable-but-incorrect outputs occurring in high-error regions. In this regime, frequency-domain packet scalar energy is the strongest predictor of reconstruction error (r=0.236, p<10−7), outperforming pixel residuals. A linked-view interface exposes heatmap overlays and timeline prioritization to support monitoring and intervention.

The use of GPS in sports through personal fitness trackers has become widespread. For amateur team sports, the decreasing cost of GPS devices is making tracking devices a more viable option. However, once the data is generated, there are still barriers in accessing tools to explore the group trajectories in ways that would allow amateur coaches or players to gain positional information with strategic relevance. This paper presents an initial investigation into the requirements of amateur football (soccer). Six coaches and players were interviewed from three different amateur football clubs in Australia. Thematic analysis was used to identify key requirements for group GPS trajectory visualization. A visual analytics prototype was designed, developed, and evaluated with the participants. Amateur football visual analysis barriers, improvements for the prototype, and future work are discussed.

Sentiment analysis plays a critical role in understanding public opinion, particularly in political discourse, where sentiments evolve in response to historical events and collective emotional dynamics. However, existing sentiment analysis research largely overlooks time-grounded sociopolitical contexts and domain-specific historical data, especially for low-resource languages such as Bengali. Moreover, there are currently very few datasets that systematically integrate temporal political events with fine-grained sentiment intensity annotations. To address this gap, we introduce a time-grounded, context-aware Bengali political sentiment dataset aligned with major sociopolitical events in Bangladesh from July 2024 to August 2025. The dataset consists of manually curated political texts, annotated with a five-level sentiment intensity scale ranging from strongly negative to strongly positive. Annotation reliability is ensured through multi-annotator agreement, contextual validation using reaction signals, and GPT-assisted verification. Building on this dataset, we perform fine-grained linguistic and sentiment-intensity analyses to examine how political sentiment evolves over time and across sociopolitical contexts. We fine-tuned the existing BanglaBERT model, which is lightweight and enhanced with an attention-refinement mechanism that enables the effective capture of time-sensitive contextual signals in a low-resource setting. The proposed mechanism achieves 93.52% accuracy and a macro-F1 score of 92.70% while maintaining low computational complexity. Finally, we integrate the model into a confidence-aware visual analytics framework that supports exploratory analysis of temporally evolving political sentiment, demonstrating the practical value of the dataset and approach.

Large Language Models (LLMs) are increasingly integrated into Natural Language Interfaces (NLIs) for visualizations. When the data underlying a visualization is available along with a visualization specification, LLMs can answer user questions directly, without needing to extract visual features from the rendered chart. In this setting, the context, tasks, and reasoning guidelines are provided to the model via an instruction prompt. We evaluate the performance of GPT models on the Visualization Literacy Assessment Test (VLAT), examining how different instruction prompts affect their responses. Thereby, the prompts incorporate various prompt engineering techniques designed to guide reasoning. Our results demonstrate the potential of LLMs to answer questions about baseline visualizations with available input data. Furthermore, we provide insights into how prompt design influences the performance in this scenario. The implementation is publicly available under https://github.com/hpicgs/lumos-visualization-llm-nli.

Quantitative assessments of cultural heritage often rely heavily on subjective commentary, lacking systematic quantitative research. They frequently encounter challenges due to the multidimensional and abstract nature of the data. Effective visual analysis methods can significantly enhance the digital preservation and public dissemination of cultural heritage. Taking the Ming Dynasty's Illustrated Catalogue of Famous Porcelains Through the Ages as a case study, this paper proposes a framework integrating a multidimensional quantitative model with an interactive visualisation system. We designed and implemented the ‘ Tianlai.China ’ four-dimensional multimodal visualisation interaction system. Through digital methods, we extracted 83 porcelain pieces from the compendium's illustrations and textual records, transforming obscure scholarly literature into comprehensible, interpretable quantitative metrics and visual representations. This facilitates cultural dissemination from basic to advanced levels, enhancing public understanding and engagement. The research endeavours to translate the catalogue's subjective assessments of porcelain uniqueness into relatively objective, quantifiable metrics. It pioneers the ‘ Porcelain Uniqueness Entropy Weighted Quantification Model, ’ establishing a five-dimensional standardised evaluation system based on shape form (38%), decoration (22%), glaze colour (25%), craftsmanship (12%), and historical context (3%). To validate the visualisation system and model, a collaborative assessment experiment involving experts, the public, and AI was designed. The core contribution lies in establishing a visual analysis paradigm for textual cultural heritage preserved solely in ancient documents without physical artefacts. This provides methodological and practical references for quantifying, visualising, and evaluating cross-modal cultural data, exploring further possibilities for data visualisation and communication in complex cultural contexts.

Trajectory data record spatiotemporal information and related attributes of moving objects. High-quality trajectory data can accurately reflect behavior patterns, providing a reliable foundation for traffic management, logistics optimization, and smart city planning. However, existing research on trajectory data quality management remains limited, lacking interactive tools for addressing quality issues and making it difficult for users to explore and resolve problems in trajectory datasets. To address this gap, we present TrajectoryCurer, an interactive visual analytics system designed for trajectory data quality management. Through expert interviews and synthesis of preprocessing operations from 30 relevant articles on trajectory visualization and management, we construct a taxonomy of 19 trajectory data quality issues across six dimensions. Based on this taxonomy, we design a multilevel visualization approach encompassing Trajectory View, Point View, Map View, and Data View, enabling users to observe and address quality issues across multiple dimensions through an integrated interface. We demonstrate the effectiveness and usability of TrajectoryCurer through two usage scenarios and expert evaluations.

Decision-makers consult multiple forecasts to account for uncertainties when forming judgments about future events. While prior works have compared unaggregated and highly-aggregated designs for displaying multiple forecasts (e.g., Multiple Forecast Visualizations versus confidence interval plots), it remains unclear how partial aggregation impacts judgment. To investigate the effect of partial aggregation, we curated three designs that partially aggregate multiple forecasts. Through two large-scale studies (Experiment 1 n = 695 and Experiment 2 n = 389) across 14 judgment-related metrics, we observed that one design (Horizon Sampled MFV) significantly enhanced participants’ ability to predict future trends, thereby reducing their surprise when confronted with the actual outcomes. Grounded in empirical evidence, we provide insights into how to design visualizations for multiple forecasts to communicate uncertainty more effectively. Specifically, since no approach excels in all metrics, we advise choosing different designs based on communication goals and prior knowledge of forecasts.

Crime analysis and visualisation systems predominantly rely on incident counts, despite extensive evidence that criminal offences vary substantially in social harm. Count-based representations can therefore obscure severity patterns and misrepresent public safety impact. In this paper, we integrate crime severity modelling into visual analytics workflows by constructing an incident-level Crime Severity Score (CSS) for Cambridge, Massachusetts, using FBI NIBRS offence labels, U.S.-adapted imprisonment-day harm weights, and proportionality adjustments inspired by the UK Office for National Statistics (ONS) CSS. Severity is treated as a first-class analytic and visual variable and evaluated across multiple formulations using logarithmic, square-root, and Box–Cox transformations. Distributional analysis reveals that Box-Cox transformed severity scores yield optimal forecasting performance while preserving the relative ordering of harm. We further assess the analytic utility of severity-aware representations through short-horizon daily and weekly crime prediction tasks using gradient-boosted tree models. Results show that temporal features account for 86% of daily predictive improvement, while severity features contribute an additional 14% as a stabilising complement. For daily prediction, severity-aware models achieve a 6.9% mean absolute error (MAE) reduction and a +28.9% R2 improvement over volume-only baselines; for weekly prediction, severity features recover 5.3% R2 lost by temporal overfitting while maintaining accuracy gains. Overall, incorporating offence severity enhances both exploratory crime visualisation and predictive modelling without replacing traditional count-based views.

Artificial intelligence (AI) and generative AI (GenAI) share close technical foundations. Their ethical challenges often overlap as a result. At the same time, GenAI creates content through its distinctive generative capabilities. These same capabilities produce serious negative consequences and uncontrolled ethical risks, including deepfakes and related harms. These developments have made GenAI ethics literacy an emerging, but the existing research on it is still quite limited in scope. This survey examines both the shared ethical concerns between AI and GenAI, and the specific ethical problems unique to GenAI technologies. It uses a knowledge-attitudes-competences perspective to describe the structure of GenAI ethics literacy. In this view, GenAI ethics literacy consists of three main dimensions, knowledge, attitudes, and competences. These dimensions receive detailed explanation through the characteristics of the metaverse environment. The analysis highlights the role of AI literacy as an essential groundwork for AI ethics literacy. Understanding of AI technology and its related ethical issues directly determines the level of AI literacy. Knowledge about both AI systems and AI ethics shapes this literacy foundation. This foundation is supported by core abilities in critical thinking, technical knowledge integration, and ethical awareness. The metaverse represents a widespread virtual environment. GenAI abilities create particular challenges for literacy within this context. Media literacy requires integration into GenAI literacy frameworks moving forward. Based on these insights, this survey delivers practical guidelines. These guidelines help users understand fundamental GenAI ethical principles and highlight the value of education on AI and GenAI ethics. Users can then apply responsible practices within metaverse environments.

Early cardiovascular disease (CVD) risk detection in people with diabetes is hindered by the fragmentation of health data, where population-scale surveys offer broad coverage but limited measurement fidelity, while clinical cohorts provide precise outcomes at a restricted scale. This work investigates how conservative dataset harmonisation, supported by visual analytics, can improve robustness and interpretability in early CVD risk modelling without relying on unsupported clinical inference. We present a visualisation-oriented harmonisation framework that integrates the Framingham Heart Study (FHS) and the Behavioral Risk Factor Surveillance System (BRFSS) by aligning variables only when semantic equivalence or epidemiologically supported proxy construction is defensible. Using established machine learning models, we compare three regimes—clinical cohort data, population survey data, and a harmonised representation—through combined quantitative metrics and comparative visual analyses of distributions, prevalence patterns, and precision–recall behaviour. Across multiple models, harmonised representations reduce precision–recall volatility by approximately 15–20% relative to survey-only models and improve average precision by 10–15% compared with cohort-only models, while avoiding artificial inflation of discrimination metrics. Comparative visualisation reveals that these gains arise from improved semantic alignment and reduced dataset-specific bias rather than increased model complexity. Our results demonstrate that visual analytics plays a critical role in diagnosing distributional mismatch, validating proxy construction, and supporting robustness-aware healthcare machine learning. The proposed framework highlights how conservative data integration, when paired with transparent visual evaluation, can yield more stable and interpretable early risk modelling under heterogeneous data constraints.

A data story typically integrates data facts from multiple perspectives and stances to construct a comprehensive and objective narrative. However, retrieving these facts demands time for data search and challenges the creator’s analytical skills. In this work, we introduce InScribe, an intelligent system that automatically performs reasoning and stance-based data facts retrieval to augment the user’s statement. Particularly, InScribe leverages an LLM-based agent to construct a retrieval tree, enabling collaborative control of its expansion between users and the agent. The interface visualizes the retrieval tree as a mind map that eases users to intuitively steer the retrieval direction and effectively engage in reasoning and analysis. We evaluate the proposed system through performance evaluation, case studies, and in-depth expert interviews. Our evaluation demonstrates that InScribe can effectively retrieve multifaceted data facts from different stances, helping users verify their statements and enhance the credibility of their stories.

Visual complexity plays a crucial role in shaping how readers trust data visualizations. While recent empirical work suggests that increased complexity reduces trust, we propose a more nuanced relationship that is nonlinear and mediated by effort. We conducted an experimental study (N=759) where participants viewed visualizations of varying complexity and rated their trust in the visual design and underlying data. Participants also reported the effort they expended reading the visualization and completed a comprehension assessment of the presented information. We found no significant relationship between visual complexity and trust in data. However, we found that the relationship between visual complexity and trust in a visualization’s design unfolds in four stages, with effort playing a distinct mediation role. At a low level of visual complexity, viewers are distracted as the effort needed to process the visualization is low; at the same time, trust decreases. At a medium level of visual complexity, more effort is required to process the visualization, leading viewers to engage with a visualization while reporting higher trust. At a higher, substantial level of complexity, even higher levels of effort result in overwhelmed viewers, and we observed a decrease in trust. At the highest level of complexity captured in this work, even more effort needed leads overwhelmed viewers to disengage, while trust in the visual design increases. We present these results as a framework showing how cognitive factors interact during visualization interpretation, revealing a nuanced relationship between visual complexity and trust.

Evolving hierarchical data describes the temporal development of software repositories, organizational structures, taxonomies in medical science, and merge trees in simulation data, to name a few. Temporal treemaps visualize these evolving trees in a static 2D layout: one dimension shows the hierarchy while the other dimension is devoted to time. Hence, a temporally evolving tree node is displayed as a band of varying width indicating the data value associated with the node. In almost all practically relevant data, these bands meander up and down, vary their width, and cross each other, which is detrimental to the readability of the visualization. Previous methods focused on minimizing the number of crossings, but neglected the meandering: these so-called wiggles are geometric deviations from a straight path. In this paper, we enhance the layout of temporal treemaps by (i) proposing an efficient algorithm for minimizing wiggles in temporal treemaps. Thus, temporal treemaps can be drawn with minimal visual clutter, which is a crucial enabler for (ii) our novel application for temporal treemaps: the interactive visual comparison of two evolving hierarchies. We apply this to tracking algorithms for merge trees, where no such comparative visualizations are available yet. This new application is enabled by (iii) a novel co-optimization of the layout of two temporal treemaps. We show the effectiveness of our approach on previous and new real-world data from a selection of different domains and demonstrate its scalability.

Data visualisations are a cornerstone of modern information communication, yet their accessibility for neurodivergent audiences remains underexplored. Autistic individuals often report distinct perceptual and cognitive experiences that may influence how they interpret visualisations. However, research into accessibility and visualisation literacy for autistic adults is limited, potentially reducing their engagement with everyday information. This is one of the first studies to investigate the data visualisation experiences of autistic adults. In an online survey, autistic (N=80) and non-autistic (N=104) adults answered quantitative questions on exposure, familiarity, and comprehension alongside qualitative open-ended questions. Quantitative results showed comparable exposure and familiarity but greater difficulty with data visualisations among autistic participants. Qualitative analysis revealed overlapping usability issues but stronger sensory discomfort among autistic participants, particularly for high-saturation colours and dense patterns. We also identify key challenges, needs, and preferences related to specific design elements. While preliminary, these findings provide evidence for extending guidelines on accessible data visualisation design, offering benefit to both autistic users and the wider population.

Ontologies contain rich information about concepts and the complex relationships between these concepts. The hierarchical relationships form the backbone structure of the ontology, and diverse types of non-hierarchical relationships (associations) connect the concepts across different parts of the hierarchy. The combination of hierarchy and associations gives rise to the analysis of data and the formulation of scientific hypotheses. However, existing tools fall short in supporting complex analytical activities, such as those involving heterogeneous associations and on different levels of granularity, especially when the ontology is large with many concepts and relationships. We design and present interactive visual approaches with domain experts, to support the exploration of such complex relationships in biomedical ontologies and provide effective visual representations for this exploration. Our approaches clearly show the ontology hierarchy and emphasise the associations on top of it.They enable the investigation of the distribution of various types of associations in the ontology hierarchy, provide flexibility to perform operations to combine or compare associations, and allow these analyses on different levels of granularity (e.g., property level, class level). We implemented these approaches in a web-based tool OntoGrid, which extends prior tool OntoPlot. Case studies conducted with domain experts demonstrate the usability of OntoGrid and highlight its effectiveness in supporting complex user tasks.

Data visualization literacy is vital for preparing students for higher education in data-driven fields, yet instructions and teaching materials for high school remain limited. To address this gap, we applied participatory design (PD) and cooperative prototyping (CP) with high school teachers and data visualization experts as equal partners. Over a period of 16 months, we conducted an interview and four iterative workshops to identify teaching challenges and co-create solutions. This process produced three prototype sets, culminating in a 17-step teaching unit on the CoTinker platform for computational thinking. The unit uses a music dataset and a developed visualization tool to engage students in collaboratively creating data visualizations through hands-on learning while introducing visualization concepts, visual storytelling, and data ethics. Our work demonstrates how PD and CP can bridge research and practice, and contributes a replicable framework for integrating data visualization into high school by use of co-designed tools grounded in real-world classrooms.

Change faithfulness metrics measure how faithfully the ground truth changes in a dynamic graph G are proportionally represented as the geometric change in the drawing D of the dynamic graph. While algorithms for optimizing faithfulness metrics for static graphs have been well studied, algorithms for optimizing change faithfulness metrics remain underexplored. In this paper, we present change-faithful graph drawing algorithms to optimize change-faithfulness metrics. Specifically, we consider three change-faithful metrics, including shape-change metrics DSC, distance-change metrics DCQ, and neighborhood-change metrics DNC, and present change-faithful graph drawing algorithms for dynamic graphs to optimize each metric with three variations: Independent Random (IR), Smart Initialization (SI), and Union Graph (UG). Extensive experiments comparing the three variations of change-faithful graph drawing algorithms optimizing each metric on dynamic graphs with various structures, such as minor change and major change, demonstrate that UG obtains much better change faithfulness metrics on graphs with minor changes (≥ 80% overlap in edge set between time steps), while IR and SI obtain better change faithfulness metrics on graphs with major changes (≤ 60% overlap in edge set between time steps).

We present a comprehensive experimental evaluation of applying stochastic gradient descent (SGD) for constrained graph drawing. We separate stress minimization from constraint satisfaction, enabling flexible choice of projection methods for enforcing constraints. In controlled experiments, we benchmark the SGD-based framework against stress-majorization baselines using stress and constraint-violation, and we show improved stress across three constraint types. We further assess two projection strategies, global and sequential, and quantify their trade-offs: inserting a global projection at each SGD iteration raises stress but consistently satisfies constraints and is independent of projection order, whereas sequential projection remains order-sensitive. Finally, we demonstrate applicability to non-Euclidean layouts, including hyperbolic and toroidal domains, demonstrating flexibility for specialized visualization tasks.