Dimensionality reduction is a powerful technique for revealing structure and potential clusters in data. However, as the axes are complex, non-linear combinations of features, they often lack semantic interpretability. Existing visual analytics (VA) methods support cluster interpretation through feature comparison and interactive exploration, but they require technical expertise and intense human effort. We present LangLasso, a novel method that complements VA approaches through interactive, natural language descriptions of clusters using large language models (LLMs). It produces human-readable descriptions that make cluster interpretation accessible to non-experts and allow integration of external contextual knowledge beyond the dataset. We systematically evaluate the reliability of these explanations and demonstrate that provides an effective first step for engaging broader audiences in cluster interpretation. The tool is available at https://langlasso.vercel.app

We present a third version of the PraK system designed around an effective text-image and image-image search model. The system integrates sub-image search options for localized context search for CLIP and image color/texture and interfaces tailored to specific task types. A user study shows the benefits of sub-image search in homogeneous video collections. We further enhance the user experience of the PraK framework by re-designing the frontend and introducing taskspecific views.

Typical projection methods such as PCA or MDS rely on mapping data onto an Euclidean space, limiting the design of resulting visualizations to lines, planes, or cubes and thus may fail to capture the intrinsic non-linear relationships within data, resulting in inefficient use of two-dimensional space. We introduce the novel projection technique -cPro-, which aligns high-dimensional data onto a circular layout. We apply gradient descent, an adaptable optimization technique to efficiently reduce a customized loss function. We use selected distance measures to reduce high data dimensionality and reveal patterns on a two-dimensional ring layout. We evaluate our approach compared to 1D and 2D MDS and discuss further use cases and potential extensions. cPro enables the design of novel visualization techniques that employ semantic distances on a circular layout.

This position paper discusses the profound impact of Large Language Models (LLMs) on semantic change, emphasizing the need for comprehensive monitoring and visualization techniques. Building on linguistic concepts, we examine the interdependency between mental and language models, highlighting how LLMs and human cognition mutually influence each other within societal contexts. We introduce three primary theories to conceptualize such influences: (T1) Recontextualization, (T2) Standardization, and (T3) Semantic Dementia, illustrating how LLMs drive, standardize, and potentially degrade language semantics. Our subsequent review categorizes methods for visualizing semantic change into frequency-based, embedding-based, and context-based techniques, being first in assessing their effectiveness in capturing linguistic evolution: Embedding-based methods are highlighted as crucial for a detailed semantic analysis, reflecting both broad trends and specific linguistic changes. We underscore the need for novel visualization tools to explain LLM-induced semantic changes, ensuring the preservation of linguistic diversity and mitigating biases, while providing essential insights for the research on semantic change visualization and the dynamic nature of language evolution in the times of LLMs.

In this work, we present a visual analytics approach designed to address the 2024 VAST Challenge Mini-Challenge 1, which focuses on detecting bias in a knowledge graph. Our solution utilizes pixel-based visualizations to explore patterns within the knowledge graph, CatchNet, which is employed to identify potential illegal fishing activities. CatchNet is constructed by FishEye analysts who aggregate open-source data, including news articles and public reports. They have recently begun incorporating knowledge extracted from these sources using advanced language models. Our method combines pixel-based visualizations with ordering techniques and sentiment analysis to uncover hidden patterns in both the news articles and the knowledge graph. Notably, our analysis reveals that news articles covering critiques and convictions of companies are subject to elevated levels of bias.

In this paper, we introduce an application for exploring the effect of linguistic features on personalized argument preferences. These individual preferences are derived by measuring the impact of linguistic features on pairwise comparisons between arguments. The insights derived from this are, in turn, useful for studies of argument quality. To conduct this research, we have developed a new pipeline that covers three major components: data collection, argument comparison labeling, and data exploration, incorporating linguistic annotations of arguments and preference data. The first component has resulted in a novel corpus consisting of minimal pairs of arguments: the comparable argument corpus. For the second component, we have developed a visual interactive labeling system that structures the annotation process of pairwise comparisons. Through these annotations, we extract patterns of argument preferences using Gaussian Process Preference Learning based on linguistic feature vectors. The corresponding, personalized models are used to identify relevant features to explain argument preferences. By training individual models for different users, we gain information that allows us to compare different user groups, identifying different argumentation preferences across groups. Each of these steps is supported by novel visual analytics dashboards, facilitating data collection and annotation steps and enabling the exploration of personal preferences.

Optimal ski route selection is a challenge based on a multitude of factors, such as the steepness, compass direction, or crowdedness. The personal preferences of every skier towards these factors require individual adaptations, which aggravate this task. Current approaches within this domain do not combine automated routing capabilities with user preferences, missing out on the possibility of integrating domain knowledge in the analysis process. We introduce SkiVis, a visual analytics application to interactively explore ski slopes and provide routing recommendations based on user preferences. In collaboration with ski guides and enthusiasts, we elicited requirements and guidelines for such an application and propose different workflows depending on the skiers' familiarity with the resort. In a case study on the resort of Ski Arlberg, we illustrate how to leverage volunteered geographic information to enable a numerical comparison between slopes. We evaluated our approach through a pair-analytics study and demonstrate how it supports skiers in discovering relevant and preference-based ski routes. Besides the tasks investigated in the study, we derive additional use cases from the interviews that showcase the further potential of SkiVis, and contribute directions for further research opportunities.

To solve VAST Mini Challenge 1, we build an interactive visualization tool that allows hypothesis testing and exploratory analysis of the data. The tool contains different visualizations for metadata and audio data analysis. To analyze the recorded bird calls, we trained a Gradient Boosting-classifier to distinguish different bird species. Our tool integrates these results and visualizes them in combination with additional data allowing the user to get context information and confirm the results.