[Paper Review] DeepVenn -- a web application for the creation of area-proportional Venn diagrams using the deep learning framework Tensorflow.js

Tim Hulsen|arXiv (Cornell University)|Sep 27, 2022

Data Visualization and Analytics69 citations

TL;DR

DeepVenn is a web app that uses TensorFlow.js to generate area-proportional Venn diagrams for up to ten sets, using lists of IDs as input and an automatic optimization of overlaps.

ABSTRACT

Motivation: The Venn diagram is one of the most popular methods to visualize the overlap and differences between data sets. It is especially useful when it is are 'area-proportional'; i.e. the sizes of the circles and the overlaps are proportional to the sizes of the data sets. There are some tools available that can generate area-proportional Venn Diagrams, but most of them are limited to two or three circles, and others are not available as a web application or accept only numbers and not lists of IDs as input. Some existing solutions also have limited accuracy because of outdated algorithms to calculate the optimal placement of the circles. The latest machine learning and deep learning frameworks can offer a solution to this problem. Results: The DeepVenn web application can create area-proportional Venn diagrams of up to ten sets. Because of an algorithm implemented with the deep learning framework Tensorflow.js, DeepVenn automatically finds the optimal solution in which the overlap between the circles corresponds to the sizes of the overlap as much as possible. The only required input is two to ten lists of IDs. Optional parameters include the main title, the subtitle, the set titles and colours of the circles and the background. The user can choose to display absolute numbers or percentages in the final diagram. The image can be saved as a PNG file by right-clicking on it and choosing "Save image as". The right side of the interface also shows the numbers and contents of all intersections. Availability: DeepVenn is available at https://www.deepvenn.com. Contact: [e-mail]

Motivation & Objective

• Visualize overlaps between data sets with area-proportional Venn diagrams.
• Support up to ten sets and input as lists of IDs rather than just numbers.
• Leverage a deep learning-based algorithm to optimize circle placement so overlaps reflect data sizes.
• Provide configurable aesthetics (titles, colors, background) and output options (absolute numbers or percentages).
• Allow saving the diagram as a PNG and display intersection contents on the interface.

Proposed method

• Implement a deep learning-based algorithm in TensorFlow.js to optimize circle placements so overlaps correspond to data sizes.
• Accept input as two to ten lists of IDs to define the sets.
• Provide optional parameters for main title, subtitle, set titles, circle colors, and background.
• Offer display modes for absolute numbers or percentages in the final diagram.
• Enable PNG export by right-clicking 'Save image as'.
• Show the numbers and contents of all intersections on the interface.

Experimental results

Research questions

• RQ1 How accurately does the DeepVenn algorithm place area-proportional circles for up to ten sets?
• RQ2 What is the maximum number of sets that DeepVenn can handle effectively?
• RQ3 What input formats does DeepVenn accept and how do these affect usability and accuracy?
• RQ4 What user customization options influence the readability and utility of the resulting diagrams?
• RQ5 How does DeepVenn present and export the computed intersection data?

Key findings

• The tool can create area-proportional Venn diagrams of up to ten sets.
• An algorithm using TensorFlow.js automatically finds an optimal placement where overlaps reflect the sizes of overlaps as much as possible.
• Input is two to ten lists of IDs, with optional parameters for title, subtitle, set titles, colors, and background.
• Users can choose to display absolute numbers or percentages in the final diagram.
• The diagram can be saved as a PNG file, and the interface shows intersection contents on the right side.