# Data2Vis: Automatic Generation of Data Visualizations Using Sequence to Sequence Recurrent Neural Networks

**Authors:** Victor Dibia, Çağatay Demiralp  
**arXiv:** [1804.03126](https://arxiv.org/abs/1804.03126)  
**Date:** 2018-04-09  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/1804.03126v3/

## Key points

- Paper: "Data2Vis: Automatic Generation of Data Visualizations Using Sequence to Sequence Recurrent Neural Networks"
- Authors: Victor Dibia and Çağatay Demiralp
- Presented at IEEE VDS'18 conference
- Addresses the challenge of creating effective visualizations for users with limited time and skills in statistics and data visualization
- Existing visualization tools require manual selection of data attributes, decision-making on transformations, and specification of mappings between variables and attributes
- Proposes Data2Vis - a neural translation model that automatically generates visualizations from given datasets
- Treats visualization generation as a sequence to sequence translation problem using Vega-Lite - a declarative language - mapping data specifications to visualization specifications
- Uses a multilayered attention-based recurrent neural network (RNN) with long short-term memory (LSTM) units trained on a corpus of visualization specifications
- Qualitative results demonstrate that the model effectively learns the vocabulary and syntax required for valid visualization specifications as well as appropriate transformations such as count, bins, and mean; it also learns common data selection patterns within visualizations
- Capable of generating visualizations comparable to manually-created ones but in significantly less time
- Offers an efficient solution for rapidly creating effective visualizations without requiring extensive statistical or data visualization expertise

## Summary

The paper "Data2Vis: Automatic Generation of Data Visualizations Using Sequence to Sequence Recurrent Neural Networks" authored by Victor Dibia and Çağatay Demiralp and presented at IEEE VDS'18 conference addresses the challenge of creating effective visualizations for users with limited time and skills in statistics and data visualization. Existing visualization tools require manual selection of data attributes, decision-making on transformations, and specification of mappings between variables and attributes. To overcome these challenges, the authors propose Data2Vis - a neural translation model that automatically generates visualizations from given datasets. The model treats visualization generation as a sequence to sequence translation problem using Vega-Lite - a declarative language - mapping data specifications to visualization specifications. A multilayered attention-based recurrent neural network (RNN) with long short-term memory (LSTM) units is trained on a corpus of visualization specifications. The qualitative results demonstrate that the model effectively learns the vocabulary and syntax required for valid visualization specifications as well as appropriate transformations such as count, bins, and mean; it also learns common data selection patterns within visualizations. Data2Vis is capable of generating visualizations comparable to manually-created ones but in significantly less time. With its potential to learn more complex visualization strategies at scale, Data2Vis offers an efficient solution for rapidly creating effective visualizations without requiring extensive statistical or data visualization expertise.