Many impactful personal decisions are made by individuals with little or no prior experience in decision making who must nevertheless navigate choices that are high stakes and complex. Unlike contexts where decision makers are experienced professionals, these individuals face unfamiliar situations in which they must weigh numerous trade-offs and preferences without clear guidance or prior knowledge. To investigate how multi-criteria decision making (MCDM) visualizations can support such individuals, we conducted two exploratory design studies: one with students making their first borrowing decisions and another with job seekers who had limited experience navigating the job market. Across both studies, participants initially focused narrowly on immediate, easily quantifiable aspects such as monthly repayment ability or yes/no filters on job attributes, while overlooking long-term consequences and more nuanced qualitative trade-offs. Our studies revealed several unstructured aspects of decision making in practice, including unclear preference hierarchies, difficulties in incorporating categorical and temporal criteria, and reliance on improvised datasets. Visualization probes encouraged participants to compare alternatives more systematically, surfacing neglected criteria and prompting more reflective reasoning. Together, these findings illustrate that personal, high-stakes decisions made in real-world contexts by individuals with limited prior experience challenge the fixed-data assumptions of MCDM and highlight new directions for visualization research. Supplementary materials are available at .
Financial literacy plays a vital role in everyday life, enabling individuals to manage spending, save for goals, and make sound financial decisions. However, for many people acquiring financial knowledge remains challenging due to the complexity of concepts and a lack of motivation. Data visualization offers a promising approach by helping learners better understand complex ideas and making learning more engaging. Despite its potential, few studies have thoroughly examined its application in financial education. This paper presents an exploratory analysis of the use of data visualization in this field. We analyze commonly used visualization tools and the dimensions of financial visualization education, develop taxonomies for key dimensions, identify their patterns and relationships. We also provide recommendations for future visualization tools in financial education. Furthermore, we propose a visualization-centric pedagogical framework for financial education. Unlike generic instructional design models, this framework is empirically grounded in our review and explicitly specifies where and how visualization techniques and tools should be embedded in the financial education process, thereby addressing the current lack of domain-specific guidance. Our findings aim to support financial educators and visualization professionals in applying effective visualization strategies, while also offering researchers a foundation for further investigation into the role of visualization in financial learning.
Missing data pose a widely known challenge for multivariate data analysis. Imputation is a mechanism for replacing missing values with plausible substitutes. Multiple imputation is often used to generate several plausible datasets for a wide range of statistical models to create robust standard errors and reduce potential bias from complete case analyses. Implementing an imputation algorithm requires several decisions about settings to generate meaningful imputed datasets. The analysis tasks can be summarized as analyzing attribute relations, analyzing the missingness mechanisms, supporting the selection of imputation parameters, and analyzing the imputation outcome. We propose a visual analytics pipeline that supports decision-making in each of the required steps. Each step is supported by novel visual encodings or existing ones that are adapted to the task. In particular, we adapt all methods to operate on heterogeneous data. Key visualizations are a 2D embedding of attribute similarities for heterogeneous data, a heatmap visualization providing a scalable overview of the effect of missingness on data distributions between all pairs of attributes (and between all original and all imputed attributes), and extended parallel coordinates that can handle both heterogeneous and missing data. Moreover, we propose a novel scheme for computing the imputation order based on an interactive predictor graph visualization. We illustrate the effectiveness of our visual encodings and the interactive visual analytics pipeline within workflows executed with domain experts from epidemiology to analyze cohort studies. The domain experts reported increased confidence in the imputation result due to improved transparency.
The constant growth in data generation, driven by technological advancement, highlights the need to organize information to extract relevant knowledge. In this context, visual representations emerge as effective tools to simplify this complex task. The automation of this process can be achieved through visualization recommendation systems. This work aims to improve the understanding of data visualization recommendations by synthesizing current literature to identify research gaps and outline initial requirements for developing prototypes and tools in this area. To achieve this, we conducted a systematic literature mapping followed by forward snowballing, covering the period from 2017 to 2025, through which we carefully selected and analyzed 89 papers on data visualization recommendations. We provide an overview of visualization recommendation systems, identifying employed techniques and categorizing studies based on different recommendation approaches. We also guide the selection of algorithms and methods for developing automatic and semiautomatic recommendation systems and present lessons learned and future research possibilities.
Mental health research increasingly considers individual differences such as personality traits and sociodemographic factors, yet existing analytical practices rely largely on statistical software, which limits exploratory analysis and hypothesis generation. From a visualization perspective, there is a lack of domain-informed design studies that investigate how interactive visual analytics can support the joint exploration of these factors. We present a visualization design study conducted in close collaboration with personality psychology experts, aimed at supporting exploratory analysis of relationships between sociodemographic, personality, and mental health data. Through an iterative co-creation process involving interviews, design workshops, and prototyping, we designed FFM-MHI Vis, an interactive visualization system integrating coordinated views such as parallel coordinates, boxplots, scatterplots, and Sankey diagrams. The system was evaluated with domain experts through a task-based study comprising six analytical tasks. Results show high perceived usability (84.92 +/- 1.94), perceived usefulness (91.42 +/- 1.65%), and ease of use (87.14 +/- 1.67%). We contribute with reflections and lessons learned from the design process, demonstrating how visualization can effectively support exploratory analysis and hypothesis generation in personality and mental health research.
Stable Diffusion is a widely used text-to-image generation model. However, its outputs are highly sensitive to hyperparameters settings and often suffer issues such as semantic drift, subject misalignment and detail loss. Traditional methods rely on manually adjusting hyperparameters to alter the attention distribution and thus improve the quality of generated images, which is time-consuming and lacks precision. Therefore, we propose an attention-guided visual diagnostic system named DAttnVis, which is designed to assist users in understanding the complex inference process of the Stable Diffusion model and optimizing its parameters. The core idea is to transform high-dimensional attention signals into comparable diagnostic representations across layers using a quantifiable metric-the Attention Concentration Index (ACI). Additionally, an anomaly detection method based on Median Absolute Deviation (MAD) is proposed to accurately identify abnormal attention layers. By linking multiple views, including UNet attention flow, diagnosis and guidance, cross-attention, and historical comparison, DAttnVis constructs a comprehensive diagnostic workflow that covers global screening, structural drilling-down, semantic tracing, and result verification. Quantitative evaluation experiments, case studies and user studies demonstrate that DAttnVis can effectively reduce trial-and-error costs and debugging burdens in the model tuning process, while improving the accuracy of anomalous structure localization and key prompt attribution.
Many real-world datasets-publications over time, project progress, and health records-can be modeled as sequences of discrete events. These event sequences often exhibit irregular temporal distributions, where events cluster together in rapid succession, interspersed with periods of inactivity. Standard timeline charts with linear time axes fail to adequately represent such data, creating cluttered regions during event clusters while leaving other areas unutilized. We introduce EventLines, a novel technique that dynamically adjusts the time scale to match the underlying event distribution, enabling more efficient use of screen space. To address the challenges of non-linear time scaling, EventLines employs the time axis's visual representation itself to communicate the varying scale. We present findings from a crowdsourced graphical perception study that examines how different time scale representations influence temporal perception.
Feature importance is a technique that helps users understand machine learning models by showing how much each feature contributed to the model's predictions. For example, it can be used for housing price prediction to explain why certain features lead to higher or lower prices. Different visualizations are used to convey feature importance to users: standard bar charts, but also advanced waterfall plots and force plots as provided by SHAP. These advanced visual representations convey more information (e.g., about the additivity property of the technique). However, this may come at the expense of the figure's simplicity. Both the trade-off between these properties and the added benefit of these advanced visualizations have yet to be formally studied. In this paper, we evaluate the effectiveness of three common SHAP visualization types for users to understand how machine learning-based prediction works in a housing price prediction scenario. Each participant answered a set of quiz questions aimed to measure their basic understanding of the feature importance (the absolute impact of features), the negative or positive impact of features, and the additivity property of feature importance. By testing whether participants understood these concepts, we assert whether the advanced visual metaphors are effective in conveying additional information beyond the standard bar chart visualization. Moreover, we study whether the effectiveness is moderated by personal characteristics, such as an individual's visual familiarity and cognitive skills. Our results from 2 user experiments comprising 546 participants in total show that, despite testing specifically for the properties that waterfall plots emphasize, bar and waterfall plots perform equally well. Force plots seemed to perform worse, and these results were independent of the skills and experience of the participants. Therefore, our findings tentatively suggest that bar charts may be a preferable choice for communicating feature importance due to their simplicity and comparable effectiveness.
Generative deep learning models are increasingly adopted for climate downscaling due to their ability to efficiently produce high-resolution outputs from coarse-resolution boundary conditions. However, unlike dynamical downscaling models, which explicitly represent physical mechanisms governing regional climate behavior, generative models remain largely opaque in terms of whether and how they reflect known or plausible regional physical processes. This lack of domain knowledge connection and understanding limits scientific trust and hinders the use of generative downscaling in risk-sensitive and decision-making situations. In this work, we present a visual analytics approach that enables domain experts to explore regional physical processes reflected in trained generative climate downscaling models by examining how spatially localized, multivariable input patterns relate to model outputs across cohorts of similar predictions.
Uniform Manifold Approximation and Projection (UMAP) has become a ubiquitous tool for high-dimensional data visualization, yet its interpretation is often hindered by the "cartographic fallacy"-a cognitive bias where the embedding layout is assumed to be a faithful map of the data's intrinsic geometry, leading users to mistake algorithmic side-effects for genuine data properties. These artifacts stem not only from stochastic optimization but also from inherent mathematical assumptions regarding simplicial approximation and metric normalization. In this work, we present an interactive study aimed at diagnosing these mechanisms. We introduce a classification system derived from a suite of 10 synthetic 3D "probe" datasets, categorizing distortions into spatial logic failures, topological loss, and metric distortion. Furthermore, we demonstrate a human-in-the-loop framework that pairs layout steering with parameter tuning to correct optimization traps and reveal topological trade-offs. This approach transforms UMAP, which is in its current form a static black box, into an explorable educational instrument, helping practitioners distinguish between genuine data features and algorithmic artifacts.
Large Language Models (LLMs) are increasingly used to support complex reasoning tasks, yet their fluent textual explanations often obscure underlying assumptions and intermediate reasoning steps, making it difficult for users to verify correctness or confidently rely on the results. Existing explanation, visualization, and knowledge-graph-based reasoning approaches primarily remain read-only or model-centric, offering limited support for direct human intervention in the reasoning process. To address this gap, we present InteractiveKG, a visual analytics system that externalizes LLM-generated reasoning as persistent, editable knowledge graphs. InteractiveKG enables users to inspect, curate, and iteratively refine reasoning by directly manipulating nodes and edges, adjusting abstraction levels, and accessing contextual explanations within a unified human-in-the-loop workflow. We evaluated InteractiveKG through a controlled user study comparing text-only LLM outputs with graph-based interactions across error-correction and verification scenarios. Results show that InteractiveKG significantly improves users' ability to identify and refine problematic reasoning, strengthens trust calibration, increases decision confidence, and enhances perceived control over the reasoning process. By transforming reasoning from transient text into a manipulable graph artifact, InteractiveKG demonstrates how transparency and user agency can be systematically integrated into LLM-assisted reasoning, highlighting the importance of interactive, user-controllable representations for trustworthy human-AI collaboration.
In the era of data-intensive scientific discovery, visualization serves as a crucial cognitive tool for researchers, while captions that align with visuals are essential for accurately conveying scientific intent. However, current scientific visualization workflows face significant challenges, including high technical barriers and semantic misalignment among user intent, visual output, and textual descriptions. To address these issues, this paper proposes SciVis-AGE, a visual analytics system based on multi-agent collaboration. Its core methodologies comprise an agent-based task decomposition and operator encapsulation approach for automatic visualization generation, and a multi-agent triangular debate mechanism for semantic alignment and caption optimization. The system effectively reduces the technical burden on domain experts and, through iterative debate among Intent Guardian, Visual Verifier, and Annotation Checker agents, ensures precise alignment of generated images and captions with user intent, visual content, and highlighted features, thereby enhancing the rigor and efficiency of scientific communication.
Complex data-driven decisions often need complex data. Existing advanced visualizations, such as parallel coordinates and treemap, can help present this information but they are generally considered overly complex to read for everyday decision makers. Prior research on training techniques that reported training time included video or interactive tools that may not be fast and easy enough for decision makers to learn at the time of use. Therefore, inspired by learning theories in psychology and preferences of decision makers, we designed a set of on-the-fly training techniques based on static views to help them quickly learn and leverage these visualizations in the decision-making process. To evaluate the effectiveness of these training techniques, we conducted an empirical study. We found some evidence that three of our training techniques could help reduce the perception of difficulty in reading at least one of the complex visualizations. However, for each visualization, these training techniques were only effective in improving accuracy or response time in some of the tasks. Finally, we identified readability challenges in these two complex visualizations that could inform the design of future training techniques.
Visual data storytelling combines data, narrative, and visualization to convey insights effectively. However, determining which aspects of a dataset to emphasize can be challenging, as different audiences may require different focal points and individuals without storytelling expertise often struggle to identify what is most relevant for each group. Moreover, different communication goals, such as persuasion, knowledge transfer, or emotional engagement, require distinct storytelling strategies. Yet, existing tools rarely support users in selecting narrative patterns that align with their intent or in generating audience-specific, context-aware stories. To address this gap, we introduce NarratorVis, a system that automates audience-aware visual data storytelling. NarratorVis allows users to specify key storytelling parameters such as target audience, purpose, knowledge depth, and desired duration. These parameters are transformed into contextual guidance for story construction. The system extracts relevant facts from tabular data using rule-based logic and generates coherent narratives with visualizations, assisted by Large Language Models (LLMs) to produce fluent and audience-adaptive text. A scoring system and editing interface support further refinement. We conducted a user study in which participants used the system to complete storytelling tasks, followed by semi-structured interviews to gather feedback on their experiences, satisfaction, and perceived usefulness. The findings indicate that NarratorVis supports users in tailoring data stories to diverse audiences and increases their confidence in presentation preparation.
As artificial intelligence (AI) becomes increasingly embedded in public-sector services, governments face a communication challenge that extends beyond disclosure: how to support citizens' interpretation and evaluation of these systems. Although initiatives such as the Australian government's AI Transparency Statements (ATS) policy aim to promote accountability, it remains unclear whether conventional text-heavy disclosures effectively support public understanding and trust. This study advances a Cognitive-Narrative theory of AI transparency communication by introducing the Cognitive-Narrative Mediation (CNM) framework, which explains how visual communication formats shape perceived trustworthiness through users' experiential processing. Drawing on cognitive load theory and narrative engagement research, perceived transparency is conceptualised as a psychologically constructed judgement associated with two experiential pathways: cognitive ease and narrative immersion. In a controlled experiment ( N = 57), we compared a text-only ATS with an equivalent disclosure supplemented by a narrative-style video explainer while maintaining informational equivalence. The video condition produced higher cognitive ease and stronger narrative immersion than the text-only format. Analyses further indicate that perceived transparency was statistically associated with these experiential responses in ways consistent with the dual-pathway structure proposed by the CNM framework, which subsequently informs evaluations of agency trustworthiness across ability, integrity and benevolence dimensions. Because the video explainer represents a bundled change in visual communication format, incorporating multimodal presentation and structured narrative guidance, the findings are interpreted as format-level evidence rather than the isolated influence of narrative features. By shifting attention from disclosure content to disclosure experience, the CNM framework offers a process-oriented account of AI transparency communication and highlights how visual communication design can shape perceived transparency and the formation of institutional trust without altering underlying policy information.
Neural network models are widely used, and visualization helps to understand the black-box behavior of these models. Current visualization methods mainly focus on neural networks trained on data with an intrinsic representation (image, text, speech) and depend on human interpretation of the data. However, generic multivariate data is the most commonly used form of data, and neural network visualization options are limited. Furthermore, current methods mainly focus on showing the final learned weights and filters. In contrast, we propose an instance-based approach and show the flow of instances through the neural network to explain model behavior. The visualization method is centered around selecting instances of interest and showing the propagation of weight and activation contribution to the final classifications. This enables users to explore and understand both global and local model behavior by inspecting varying groups of instances. Combined automated and interaction techniques enable tracing importance-scored paths to explore and understand feature importance. The effectiveness of the visualization method is shown through examples and use cases on real-world classification datasets and compared with insights from computational explainability methods. Additionally, a qualitative user study confirms the effectiveness and value in analyzing neural networks using our instance-based visualization approach.
Scientific visualization pipelines encode domain-specific procedural knowledge with strict execution dependencies, making their construction sensitive to missing stages, incorrect operator usage, or improper ordering. Thus, generating executable scientific visualization pipelines from natural-language descriptions remains challenging for large language models, particularly in web-based environments where visualization authoring relies on explicit code-level pipeline assembly. In this work, we investigate the reliability of LLM-based scientific visualization pipeline generation, focusing on vtk.js as a representative web-based visualization library. We propose a structure-aware retrieval-augmented generation workflow that provides pipeline-aligned vtk.js code examples as contextual guidance, supporting correct module selection, parameter configuration, and execution order. We evaluate the proposed workflow across multiple multi-stage scientific visualization tasks and LLMs, measuring reliability in terms of pipeline executability and human correction effort. To this end, we introduce correction cost as metric for the amount of manual intervention required to obtain a valid pipeline. Our results show that structured, domain-specific context substantially improves pipeline executability and reduces correction cost. We additionally provide an interactive analysis interface to support human-in-the-loop inspection and systematic evaluation of generated visualization pipelines.
The importance of data-driven decision-making is rapidly increasing thanks in part to the growing availability and accessibility of data sets and analysis tools. Yet, applicable insight can be difficult due to biases and anomalies in data. An often overlooked phenomenon is mix effects, in which subgroups of data exhibit patterns opposite to the data as a whole. This phenomenon is widespread and often leads inexperienced analysts to draw incorrect statistical conclusions. In this paper we present Wiggum, an interactive visual analysis system for uncovering both mix effects and special cases known as Simpson's paradox. A Python-based web implementation of Wiggum lets users interactively analyze multidimensional data sets to reveal various forms of mix effects. Through use cases, we describe how Wiggum supports the examination of mix effects in three real data sets and demonstrate how a combination of visualization techniques-heatmaps, trend plots, small multiples, coordinated multiple views, and dynamic queries for multi-attribute drill-down-are effective for analyzing mix effects. We conducted a user study to evaluate Wiggum, focusing on users' ability to comprehend the statistical concepts, identify the corresponding visual patterns, and perform common analysis tasks correctly and efficiently. We discuss usability issues, utility limitations, and outline future directions to improve Wiggum.
Comparing graphs to identify similarities is a fundamental task in visual analytics of graph data. To support this, visual analytics systems frequently employ quantitative computational measures to provide automated guidance. However, it remains unclear how well these measures align with subjective human visual perception, thereby offering recommendations that conflict with analysts' intuitive judgments, potentially leading to confusion rather than reducing cognitive load. Multimodal Large Language Models (MLLMs), capable of visually interpreting graphs and explaining their reasoning in natural language, have emerged as a potential alternative to address this challenge. This paper bridges the gap between human and machine assessment of graph similarity through 3 interconnected experiments using a dataset of 1881 node-link diagrams. Experiment 1 collects relative similarity judgments and rationales from 32 human participants, revealing consensus on graph similarity while prioritizing global shapes and edge densities over exact topological details. Experiment 2 benchmarks 16 computational measures against these human judgments, identifying Portrait divergence as the best-performing metric, though with only moderate alignment. Experiment 3 evaluates the potential of 3 state-of-the-art MLLMs (GPT-5, Gemini 2.5 Pro, Claude Sonnet 4.5) as perceptual proxies. The results demonstrate that MLLMs, particularly GPT-5, significantly outperform traditional measures in aligning with human graph similarity perception and provide interpretable rationales for their decisions, whereas Claude Sonnet 4.5 shows the best computational efficiency. Our findings suggest that MLLMs hold significant promise not only as effective, explainable proxies for human perception but also as intelligent guides that can uncover subtle nuances that might be overlooked by human analysts in visual analytics systems.
