New Visualization Framework Explains Why Data Analysis Results Happen

Data analysts often receive the outputs of their work—cluster labels, outlier flags, classification categories—stripped of the context needed to explain them. A result table may say that certain regions belong to Cluster 0 and others to Cluster 1, but it rarely shows why. A new open-access paper published in Data Science and Engineering on 6 May 2026 introduces CAMV, a framework for Context-Aware Multi-View Visualization, designed to close exactly that gap. Instead of displaying analytical results in isolation, CAMV automatically extracts relevant contextual information from the underlying dataset and generates coordinated multi-view visualizations that place each result alongside the evidence that may explain it.

The framework was developed by Yanna Lin, Liwenhan Xie, Leixian Shen, Zhuochen Jin, Sicheng Song, Zikun Deng, Huamin Qu and Ke Xu, working across academic and industry affiliations. According to the authors, existing tools tend to visualize results alone, offering limited access to the associated contexts needed to understand why those results occur. CAMV consists of two major components: one that identifies key data subspaces and fact types as explanatory context, and another that generates coordinated multi-view visualizations of both the analysis results and their contextual information. The team built an interactive prototype and evaluated it through performance tests, a use case, a comparison study with 12 participants using GPT-4o as a baseline, and a qualitative comparison.

The technical core of CAMV is its analytical intent inference module. The system takes as input a data table and its associated analysis results, organized as a structured JSON object that captures the analysis pipeline: the aggregation attributes that set the level of analysis, the metrics of interest, the algorithm applied, and the resulting outputs. Working from a pipeline identified through collaboration with seven analysis experts at a leading company’s analysis department, the framework first enriches the result table with contextual features derived from the original fact table. Categorical and temporal attributes are summarized with distinct-value counts and distributions, while numerical attributes receive basic statistics such as averages, sums, minima, maxima and histogram-based distributions, all aggregated at the same analysis level as the original result.

Once the augmented table is built, CAMV scores every contextual feature along three dimensions. A sensitivity score measures how much a feature influences the analytical result by re-running the algorithm on that feature alone and comparing the output with the original result, using the Adjusted Rand Index for clustering tasks and Jaccard distance for subset-based outputs such as outlier detection. A divergence score, computed with Jensen–Shannon divergence, quantifies how differently the feature is distributed across result categories. A correlation score captures the statistical relationship between the feature and the metrics used in the analysis. In the implementation, the three scores are combined with equal weights, a choice the authors say was informed by industry analysts who regard these perspectives as equally important signals. The top-ranked features are then assembled into data subspaces, each linked to a data fact type such as distribution divergence or feature–metric relationship divergence, and ranked by silhouette score, which rewards both separation between result groups and cohesion within them.

The visualization module maps each selected subspace to an appropriate chart using rule tables informed by data fact type, data type and data role. The prototype presents result visualizations and contextual visualizations in two coordinated panels, with hover highlighting, click selection and lasso interaction linking data elements across views. The authors report that the pipeline scales linearly with dataset size and attribute count: on average, processing time increased by roughly one second per additional 20,000 rows, per 16 additional target attributes, or per 8 additional other attributes, with a dataset of 20,000 rows and 9 total attributes taking around one second to process.

To illustrate the framework in practice, the paper walks through a use case on the Missing Migrants dataset from Kaggle, which comprises more than 13,000 rows and 19 columns documenting migrant incidents. An analyst aggregated the data by Region of Origin and applied KMeans clustering based on mortality-related attributes, producing three clusters. Although the clusters separated clearly in a two-dimensional principal component projection, the abstract components left the analyst unsure what they meant. CAMV’s contextual views then showed that regions in Cluster 0 had the highest numbers of dead and missing individuals and the lowest survival rates, along with high male migrant rates, incidents spread across all twelve months, and diverse causes of death. Cluster 1 regions showed the opposite profile, with incidents concentrated in a single month and homogeneous causes. Clicking an anomalous bar in a histogram revealed that one Cluster 1 region with an unusually high male rate was Oceania, tied to a single tragic event, confirming that the cluster was defined by temporal concentration rather than male participation alone.

The user study provided quantitative support. Twelve participants, graduate students and senior engineers aged 22 to 33 with high self-reported familiarity with data analysis, interpreted clustering and outlier detection results on the same dataset using both CAMV and ChatGPT with GPT-4o as the API, in counterbalanced order. Participants using CAMV required significantly less time to understand the analytical results than those using ChatGPT, with mean completion times of 355.0 seconds versus 887.9 seconds. On average, participants reviewed 6.25 of the 10 visualizations provided to reach a thorough understanding, and normalized precision@K remained relatively stable at means of 0.94, 0.86 and 0.81 for K equal to 3, 6 and 10, which the authors interpret as roughly 87 percent of recommended visualizations being considered helpful. Questionnaire ratings on a five-point Likert scale showed CAMV significantly outperforming GPT-4o across all measured dimensions, with all comparisons reaching statistical significance at p below 0.01.

Qualitative feedback highlighted why the coordinated views mattered. Participants praised CAMV’s non-linear, information-dense and data-grounded presentation, noting that they could see everything at once rather than extracting information linearly through prompts. All participants expressed high trust in CAMV because its reasoning was grounded in visible data, whereas ChatGPT was criticized for unverifiable claims, a lack of direct data traceability, and verbose, unstructured responses. ChatGPT was still valued for its natural-language narration and external knowledge, and several participants expressed interest in combining the two approaches, for example by adding lightweight summaries alongside CAMV’s charts. In the authors’ qualitative comparison, they observed that when users prompted GPT-4o to explain clustering results, the model often merely summarized the result table rather than proactively exploring the original data for underlying reasons, forcing users into tedious iterative prompting.

The authors are careful about the framework’s boundaries. CAMV currently supports analysis tasks with discrete output labels—clustering, binary and multi-class classification, outlier detection and discretization-based binning—because it identifies important attributes through distributional divergence across result groups. It does not support tasks with continuous outcomes such as regression, which the authors note pose significant challenges for contrastive explanation. The current implementation covers five representative chart types and two intent types, relies on predefined heuristic mapping rules, and has not formally tested the sensitivity of parameters such as the number of top features and equal scoring weights, since evaluating interpretability changes requires substantial human studies. The system also lacks advanced interactions such as drill-down across coordinated views. Because the project is funded by a company, the system code cannot be released, although the user study data and demonstration materials will be made publicly available. The authors hope the work emphasizes the importance of efficient, multi-perspective, context-aware information access for interpreting analytical results, and suggest future integration with large language models to combine external knowledge and natural-language narratives with data-grounded visual evidence.

Subject of Research: A context-aware multi-view visualization framework that automatically infers explanatory data context for interpreting analytical results.

Article Title: CAMV: A Framework for Context-Aware Multi-View Visualization of Data Analysis Results

Article References: Lin, Y., Xie, L., Shen, L., Jin, Z., Song, S., Deng, Z., Qu, H., & Xu, K. (2026). CAMV: A Framework for Context-Aware Multi-View Visualization of Data Analysis Results. Data Science and Engineering. https://doi.org/10.1007/s41019-026-00353-x

Image Credits: AI Generated

DOI: 10.1007/s41019-026-00353-x

Keywords: data visualization, multi-view visualization, visual analytics, explainability, clustering, outlier detection, large language models, GPT-4o, user study, data analysis, CAMV, interactive dashboards