Pie Chart or Pizza: Identifying Chart Types and Their Virality on Twitter

Pavlos Vougiouklis, Leslie Carr, Elena SimperlICWSM 2020 · OpenReview

Abstract: We aim to understand how data, rendered visually as charts or infographics, “travels” on social media. To do so we propose a neural network architecture that is trained to distinguish among different types of charts, for instance line graphs or scatter plots, and predict how much they will be shared. This poses significant challenges because of the varying format and quality of the charts that are posted, and the limitations in existing training data. To start with, our proposed system outperforms related work in chart type classification on the ReVision corpus. Furthermore, we use crowdsourcing to build a new corpus, more suitable to our aims, consisting of chart images shared by data journalists on Twitter. We evaluate our system on the second corpus with respect to both chart identification and virality prediction, with promising results.

The Challenge of "Messy" Data

While data visualizations are a staple of modern communication, most research into automated chart classification has relied on "idealized" datasets—clean, standalone images of bar charts or scatter plots. In the real world, specifically on social media platforms like Twitter, charts are rarely that tidy. They are often embedded in complex infographics, overlaid with text, or surrounded by "chartjunk" (decorative elements).

Examples of standard chart types (top) vs. real-world, "messy" charts found on social media (bottom).

The authors identify a significant gap: existing models trained on clean data fail to generalize to the chaotic visual landscape of Twitter. Furthermore, there has been little effort to quantify the "virality" of these charts. Why does one line graph get thousands of retweets while another goes unnoticed? This paper tackles both the identification of charts in the wild and the prediction of their social media impact.

Methods: A Multi-Modal Approach

The authors proposed a two-part pipeline to address these challenges:

  1. Chart Identification (ConvNet): They adapted the VGGNet architecture into a more efficient Convolutional Neural Network (ConvNet). Crucially, they reduced the number of learnable parameters from roughly 130 million to just 2.5 million. This prevents "overfitting," where a model memorizes a small dataset rather than learning general patterns. This network classifies images into 11 chart types or a "non-chart" category.
  2. Virality Prediction (Multi-modal Fusion): To predict retweets and likes, the authors built a multi-modal architecture that processes three distinct signals: visual features from the chart image, textual features from the tweet, and social features from the author's profile (e.g., follower count).
The multi-modal fusion architecture combining textual, social (author), and visual features to predict chart virality.

The system projects these three streams into a shared feature space to perform a regression task, predicting the natural logarithm of retweet and like counts.

Datasets and Insights

The study utilized three primary datasets: ReVision+ (an augmented benchmark), DataTweet (34,491 tweets from major data journalism accounts), and DataTweet+ (3,000 manually annotated images from Twitter).

A key insight was that models trained on idealized academic datasets saw a massive performance drop when tested on Twitter data. Fine-tuning on the "messy" DataTweet+ set improved accuracy by over 15%, proving that "noise" (text overlays and logos) is a feature the model must learn to navigate. Furthermore, the best predictions for virality came from combining all three modalities. While an author's "clout" is a strong predictor, visual cues like line density and color coding contribute independently to how much a chart is shared.

Key Results

  • Classification: The system achieved a 93.61% test accuracy on the 12-class ReVision+ task, outperforming previous state-of-the-art models.
  • Binary Detection: When simply asked "is this a chart or not?", the model achieved an F1 score of 98.57%.
  • Virality: In the virality regression task, the multi-modal approach achieved the lowest error rates (0.858 RMSE for likes and 0.962 for retweets).

Limitations and Future Work

The authors acknowledge that the current model struggles with "dashboards" containing multiple chart types and extremely stylized infographics. Future work could incorporate temporal dynamics—how retweets accumulate over time—and explore generative applications to help users design charts specifically optimized for social media engagement.

Summary updated

Abstract

We aim to understand how data, rendered visually as charts or infographics, “travels” on social media. To do so we propose a neural network architecture that is trained to distinguish among different types of charts, for instance line graphs or scatter plots, and predict how much they will be shared. This poses significant challenges because of the varying format and quality of the charts that are posted, and the limitations in existing training data. To start with, our proposed system outperforms related work in chart type classification on the ReVision corpus. Furthermore, we use crowdsourcing to build a new corpus, more suitable to our aims, consisting of chart images shared by data journalists on Twitter. We evaluate our system on the second corpus with respect to both chart identification and virality prediction, with promising results.
BibTeX
@inproceedings{vougiouklis2020pie,
  author = {Pavlos Vougiouklis and Leslie Carr and Elena Simperl},
  title = {{Pie Chart or Pizza: Identifying Chart Types and Their Virality on Twitter}},
  booktitle = {ICWSM},
  year = {2020},
  url = {https://openreview.net/forum?id=0f_Wk8qymz_},
}
Figure