Let's dig into the data (from China)bookmark_border

Graph visualization is an essential way to understand the relationships within graph-structured data, and node-link diagrams are widely adopted because of their intuitive representation. However, generating satisfactory layouts remains challenging, because multiple valid layouts may exist for a single graph, and their usefulness often depends on users' analytical goals. Conventional layout algorithms (e.g., force-directed methods) typically require users to tune low-level parameters to achieve desirable results, which can be time-consuming and non-intuitive for users. Recent deep learning-based methods largely operate automatically but provide limited support for expressing user intentions. We propose DiffGraph, a controllable graph layout generation method with diffusion models. DiffGraph leverages a denoising diffusion probabilistic model, with graph neural networks serving as its core architecture. Instead of requiring users to tune low-level algorithmic parameters, DiffGraph allows users to provide high-level guidance, such as relative node positions or desired global layout properties. This design enables users to incorporate domain knowledge, highlight regions of interest, or enforce specific structural patterns, thereby supporting diverse analytical needs. Qualitative and quantitative evaluations show that DiffGraph can generate high-quality and diverse layouts while effectively incorporating user-defined constraints. A user study further demonstrates that participants generally understood the constraint mechanism and perceived the generated layouts as consistent with their intentions, while an extended usage scenario illustrates how DiffGraph can support iterative and practical graph exploration.

Group discussion-based teaching is widely used to foster collaborative learning, yet teachers in physical classrooms often struggle to simultaneously monitor multiple groups and quickly diagnose a target group before intervening. Existing visual analytics tools primarily support post-hoc analysis on desktop, providing limited support for in-situ walk-around teaching. To address this gap, we present MobileGroupVis, a mobile visual analytics system for in-situ analysis of classroom group discussions. MobileGroupVis integrates multi-group monitoring, single-group diagnosis, and instructional intervention into a concise analytical workflow tailored for small-screen touch interaction. The system is powered by a lightweight streaming analysis pipeline that converts group audio into structured discussion data and further extracts interaction patterns, topic progression, and topic deviation through a dialogue analysis module. To enable both glanceable overview and traceable diagnosis, we design six coordinated views, including a compact glyph that visually encodes word count, interaction intensity, and topic deviation for efficient cross-group comparison and anomaly localization, along with detailed views for opinion evolution, interaction dynamics, topic coverage, and dialogue records. We evaluate MobileGroupVis through two case studies and expert interviews. The results provide preliminary evidence that MobileGroupVis supports teachers in understanding discussion processes, identifying groups in need of attention, and facilitating in-class intervention.

Information distortion is inevitable in the dimensionality reduction of high-dimensional data. To help understand and evaluate such distortions, researchers have proposed various quality measures for projections. Among them, measures based on quantifying neighborhood preservation are widely adopted. However, these measures do not always align with human visual perception. For instance, a projection may achieve a neighborhood preservation score above 0.99, while its k-nearest neighbors in the high-dimensional space are mapped beyond the 5k-neighborhood in the projection space. Such misalignment between perceived projection quality and quantitative measures can severely mislead users' interpretation of the resulting projections. To bridge this gap, we revisit projection quality assessment through the lens of visual perception. Our work unfolds in four key steps: First, we introduce a neighborhood-based labeling scheme for projection quality, establishing a reliable foundation for perception-driven evaluation. Second, we design a carefully curated sample set paired with a pairwise contrast-based labeling strategy, which significantly reduces annotation effort while maintaining effectiveness. Third, leveraging these labeled samples, we train a deep neural network to learn and simulate human visual perception of neighborhood preservation, resulting in a novel perception-driven quality measure. Finally, armed with this perception-driven measure, we conduct a systematic comparative evaluation of existing neighborhood quantification-based measures, providing perception-grounded insights into them.

Multimodal large language models (MLLMs) are increasingly expected to automate visualization development by generating code directly from visual designs. However, existing evaluations focus on single-chart generation, overlooking coordinated multi-view interface construction, a practically demanding task that requires jointly managing data semantics, multi-view coordination, and interactive logic. As a result, MLLMs' capabilities in this area remain largely unexamined, while the field also lacks a dedicated benchmark for systematic assessment. To address this gap, we introduce MV-Bench, the first benchmark for evaluating MLLMs on coordinated multi-view interface construction. Rather than relying on incomplete or inconsistent open-source implementations, we leverage Tableau workbook files as a principled source of ground-truth, as they explicitly encode data bindings, visual mappings, and interactions. We develop a multi-stage pipeline that converts these specifications into executable web-based interfaces via structured intermediate representations. Using this pipeline, we construct 92 base interfaces and systematically recombine them across chart types, datasets, and interaction patterns, producing 1,048 verified benchmark instances with executable code, rendered interfaces, datasets, and interaction annotations. Using MV-Bench, we evaluate five state-of-the-art MLLMs, queried under a single-pass setting, across metrics measuring visual fidelity, data binding correctness, and interaction completeness. While the strongest model achieves 75.45% accuracy in visual layout reproduction, performance drops sharply to 21.71% on data binding and 11.68% on interaction completeness. These results demonstrate that current MLLMs can reproduce visual appearance but remain fundamentally limited in generating the data semantics and interactive logic required by coordinated multi-view interfaces. Further experiments with iterative refinement improve code executability but do not substantially narrow the gap in data binding and interaction generation.

Air quality forecasting models increasingly incorporate physical priors, but it remains unclear whether their learned representations encode physical knowledge that can be recovered and, if so, to what extent. Answering this question is challenging because air quality forecasting is highly heterogeneous across time, space, and environmental conditions, causing the physical knowledge to become uneven or fragile in specific scenarios. We present PhysProbeVA, a probe-driven visual analytics framework for quantifying, diagnosing, and enhancing physical knowledge in air quality forecasting models. PhysProbeVA employs probing tasks around two core transport mechanisms, advection and diffusion, and uses selectivity-based measurements to estimate the recoverability of physical knowledge from frozen hidden representations. It further integrates the resulting evidence into coordinated visualizations that support progressive analysis from global assessment to regime localization, spatial diagnosis, subgroup slicing, and curriculum construction for targeted model refinement. Through a case study, a user study with 12 participants, and quantitative enhancement experiments, we show that PhysProbeVA helps users identify weak-knowledge scenarios, transform them into actionable curriculum candidates, and improve both forecasting performance and representation-level physical knowledge in the refined models.

Force-Directed Placement (FDP) is a widely used approach for network visualization, yet scaling it to massive graphs while preserving clear community structures remains a major computational and visual challenge. Existing approximation methods often rely on auxiliary data structures (e.g., spatial trees), which introduce substantial memory overhead; furthermore, traditional power-function-based forces frequently fail to separate dense clusters effectively. In this paper, we present a negative sampling-based algorithm that achieves O(|E|) time complexity with a low memory footprint, without requiring complex multi-level representations. In a first step, we introduce a linearly normalized degree-weighting scheme, which, combined with short-range bounded $t$-distribution forces, effectively untangles dense structures and enhances visual cluster separation. To optimize for this formulation efficiently, we introduce an edge-centric negative sampling strategy that naturally reconstructs the global degree-weighted objective. Furthermore, we design a lock-free, bundle-based parallelization scheme that leverages the sparsity of stochastic updates to achieve significant speedups while mitigating access conflicts. Comprehensive evaluations on 12 large-scale graphs demonstrate that the proposed method outperforms state-of-the-art algorithms in neighborhood preservation and cluster separation. Compared to existing baselines, our method reduces memory consumption by 72% on average and leverages simple GPU parallelism to generate a high-quality layout for a graph with 4 million nodes and 34 million edges in below 10 seconds.

Topic modeling is a fundamental technique for analyzing large-scale texts. However, its abstract and static outputs often impose a substantial cognitive burden on analysts. To bridge this gap, we propose EvoInsight, an LLM-based conversational visual analytics framework that transforms abstract topic evolution into dynamic data videos, termed Topic-Oriented Videos (TOVs). EvoInsight overcomes two primary limitations of existing data video techniques. First, existing tools rely on pre-defined narrative scripts, making them ill-suited for exploring unknown corpora. To address this, EvoInsight integrates an LLM-driven method that interactively discovers latent topics from a vectorized knowledge base, extracts relevant events, and organizes them into a coherent narrative backbone. Second, existing visualization techniques struggle to convey the complex, multi-entity relationships inherent in topic evolution. Therefore, our framework employs an end-to-end visual generation pipeline featuring multi-agent collaboration and image-to-image enhancement. This ensures semantic faithfulness and stylistic coherence across the generated video frames. Comprehensive evaluations, including a quantitative user study and qualitative domain-expert interviews, demonstrate that EvoInsight significantly outperforms ablations in narrative fluency and visual consistency, achieving quality comparable to human-crafted videos.