Insight-Guided Visualization Recommendation

Bin Chen, Yuhan Zhang, Ruizhen HuJournal of Computer-Aided Design & Computer Graphics 2023 · DOI

Abstract

Exploration and analysis of high-dimensional datasets require extracting valuable insights and generating proper charts from data. To lower the barriers of creating effective visualizations, a new insight-centric visualization recommendation framework is presented that recommends both what to visualize (data query recommendations) and how to visualize it (visual encoding recommendations). For the former, instead of using single indicators, multiple analysis methods and learning-based algorithms are utilized to automatically rank data queries based on the insights and the most informative ones are provided to the users. For the later, rule-based and learning-based methods are combined to generate visual mappings from data queries to charts and further highlight the key information in the generated charts to enhance their readability. It is demonstrated that proposed method achieves higher accuracy than existing methods for the chart type prediction task with real-world charts dataset. A user study is also conducted to show that proposed method can help users quickly explore new data tables and find important insights.

Introduction

Exploratory Data Analysis (EDA) is often a high-barrier task. Users must navigate complex data structures, select relevant dimensions, apply transformations like aggregation, and choose appropriate visual encodings. Existing recommendation systems frequently suffer from "single-indicator bias"—relying on a single metric like Pearson correlation—or lack explainability by using "black-box" deep learning models that struggle with sparse features.

This research, published in the Journal of Computer-Aided Design & Computer Graphics (2023), proposes an insight-centric framework to automate the selection of both data queries and visual encodings. By combining multi-algorithm statistical mining with machine learning, the system identifies significant patterns and recommends accurate chart types.

Methodology

The framework follows a three-stage pipeline designed to bridge the gap between raw data and visual storytelling.

The system dashboard showing the three-layered organization of insights (1D, 2D, 3D) for a COVID-19 dataset.

Data Query Enumeration and Layering

The system organizes data fragments by dimensionality (1D, 2D, 3D) and automatically applies transformations such as grouping and aggregation. This layering mimics human cognitive patterns, starting with simple views and progressing to complex ones, while preventing repetitive data presentation across categories.

Multi-Algorithm Insight Mining

Instead of a single metric, the system employs a battery of tests for three insight types:

  • Outliers: Evaluated via IQR, Z-score, LOF, and Isolation Forest.
  • Correlation: Measured using Pearson, Spearman, and Mutual Information.
  • Distribution: Analyzed through Skewness, Kurtosis, and KS tests. Scores from these diverse algorithms are normalized and aggregated to rank the most informative data fragments.

Visual Encoding and Highlighting

The system uses calculated insight scores as features for a Decision Tree (DT) classifier to predict chart types. This approach proved more interpretable and accurate than previous neural network approaches. Furthermore, the framework automatically annotates charts to guide users toward discovered insights.

An example of automated highlighting where local peaks in daily new deaths are marked with red dots to guide user attention.

Datasets and Evaluation

The model was trained on a cleaned subset of the VizML Dataset, consisting of 120,330 high-quality data-chart pairs. For user validation, real-world datasets including COVID-19 statistics, Hollywood movie profits, and Olympic medal counts were utilized.

Key Results

The Decision Tree model achieved exceptional accuracy: 97.53% for 1D/2D charts and 99.98% for 3D charts, significantly outperforming the VizML baseline. The system is also highly efficient, processing 1D insights in under one second on standard hardware. A user study with 15 participants showed a 79.4% accuracy rate in answering complex data questions, with positive feedback on the tool's ability to facilitate rapid exploration of unfamiliar data.

Conclusion and Future Work

While effective, the current system focuses on global dimensions and may miss insights within specific data "slices" (sub-spaces). Future work aims to expand chart diversity to include heatmaps and parallel coordinates, and to optimize color palettes for better insight highlighting.

Summary updated

Abstract

<p indent="0mm">Exploration and analysis of high-dimensional datasets require extracting valuable insights and generating proper charts from data. To lower the barriers of creating effective visualizations, a new insight-centric visualization recommendation framework is presented that recommends both what to visualize (data query recommendations) and how to visualize it (visual encoding recommendations). For the former, instead of using single indicators, multiple analysis methods and learning-based algorithms are utilized to automatically rank data queries based on the insights and the most informative ones are provided to the users. For the later, rule-based and learning-based methods are combined to generate visual mappings from data queries to charts and further highlight the key information in the generated charts to enhance their readability. It is demonstrated that proposed method achieves higher accuracy than existing methods for the chart type prediction task with real-world charts dataset. A user study is also conducted to show that proposed method can help users quickly explore new data tables and find important insights.
BibTeX
@article{chen2023insight,
  author = {Bin Chen and Yuhan Zhang and Ruizhen Hu},
  title = {{Insight-Guided Visualization Recommendation}},
  journal = {Journal of Computer-Aided Design \& Computer Graphics},
  year = {2023},
  doi = {10.3724/sp.j.1089.2023.19312},
}
Figure