# KG4Vis: A Knowledge Graph-Based Approach for Visualization Recommendation

**Authors:** Haotian Li, Yong Wang, Songheng Zhang, Yangqiu Song, Huamin Qu  
**arXiv:** [2107.12548](https://arxiv.org/abs/2107.12548)  
**Date:** 2021-07-27  
**Summary page:** https://www.summarizepaper.com/en/paper/2107.12548v3/

## Key points

- KG4Vis is a knowledge graph-based approach for visualization recommendation.
- The goal is to lower the barriers for general users without a background in data visualizations to create effective visualizations.
- Existing rule-based approaches require manual specifications of visualization rules by experts, while machine learning-based approaches lack explainability.
- KG4Vis addresses these limitations by leveraging a knowledge graph framework that models the mapping rules between data and visualizations.
- The knowledge graph consists of three types of entities: data features, data columns, and visualization design choices.
- These entities are connected through relations to represent the relationships between them.
- A TransE-based embedding technique is employed to learn the embeddings of entities and relations using existing dataset-visualization pairs.
- These embeddings capture the desirable visualization rules.
- With the learned embeddings, KG4Vis can infer effective visualizations from a new dataset using semantically meaningful rules derived from the knowledge graph.
- The approach is evaluated through quantitative comparisons, case studies, and expert interviews to demonstrate its effectiveness.
- KG4Vis offers an automated visualization recommendation solution that does not require manual specification of rules and provides good explainability.
- It has the potential to enable general users to create effective visualizations more easily and facilitate wider adoption of visualization approaches.

## Summary

The paper introduces KG4Vis, a knowledge graph-based approach for visualization recommendation. The goal is to lower the barriers for general users without a background in data visualizations to create effective visualizations. Existing rule-based approaches require manual specifications of visualization rules by experts, while machine learning-based approaches lack explainability. KG4Vis addresses these limitations by leveraging a knowledge graph framework that models the mapping rules between data and visualizations. The knowledge graph consists of three types of entities: data features, data columns, and visualization design choices. These entities are connected through relations to represent the relationships between them. To learn the embeddings of entities and relations, a TransE-based embedding technique is employed using existing dataset-visualization pairs. These embeddings capture the desirable visualization rules. With the learned embeddings, KG4Vis can infer effective visualizations from a new dataset using semantically meaningful rules derived from the knowledge graph. The approach is evaluated through quantitative comparisons, case studies, and expert interviews to demonstrate its effectiveness. Overall, KG4Vis offers an automated visualization recommendation solution that does not require manual specification of rules and provides good explainability. It has the potential to enable general users to create effective visualizations more easily and facilitate wider adoption of visualization approaches.