Popularity Prediction of Online Petitions using a Multimodal DeepRegression Model

Kotaro Kitayama, Shivashankar Subramanian, Timothy BaldwinALTA 2020

Introduction

Predicting the success of online petitions is a significant "cold start" challenge. While most existing research relies on early signature signals to forecast final counts, this approach does not assist organizers in optimizing content before publication. In the attention economy of platforms like Avaaz.org, a petition's visual presentation may be as critical as its written plea. This research addresses the "at-submission" prediction problem by developing a multimodal approach that integrates text, images, and metadata to predict popularity across multiple languages.

Methodology

The researchers framed popularity prediction as a regression task, aiming to predict the log-transformed number of signatures. They evaluated three distinct model architectures to determine the impact of different modalities:

  1. Text-only: A comparison between traditional CNN-based regressors and BERT-based models. For multilingual tasks, mBERT (multilingual BERT) was employed.
  2. Image-only: Visual features were extracted from the petition's primary image using a pre-trained Inception-ResNet v2 model.
  3. Combined Model: This late-fusion architecture freezes the parameters of the text and image extractors, feeding their outputs into a "Joint Embedding Network." This network utilizes fully connected layers and L2 normalization to create a shared representation, which is then combined with author metadata (country of origin).
Architectural diagrams for (a) text-only, (b) image-only, and (c) the proposed combined multimodal model.

A core technical innovation is the introduction of a joint loss function. By combining Mean Squared Logarithmic Error (MSLE) with Mean Absolute Percentage Error (MAPE), the model is forced to prioritize the relative deviation between predicted and actual signatures, aligning the optimization process more closely with the goal of popularity prediction.

Dataset and Multilingual Scope

The study utilized a dataset of approximately 75,000 petitions from Avaaz.org, spanning English, Portuguese, French, and Spanish. A unique contribution of this work was the crawling of image content for these petitions to create a new multimodal benchmark. Notably, the researchers observed that 45% of petitions used platform-provided "default" images, which added complexity to the visual analysis task.

Results and Discussion

The models were evaluated using MAPE (lower is better) and Spearman’s rank correlation ($\rho$, higher is better). The results consistently demonstrated that while text remains the primary driver of prediction accuracy, the inclusion of images and metadata provides essential context that improves performance.

Performance comparison across different modalities and loss functions, showing the benefit of the joint MSLE + MAPE approach.

Key findings include:

  • Multimodality is Superior: In all experimental settings, the combined model outperformed single-modality baselines. For monolingual English petitions, the combined model achieved a MAPE of 34.9% and a Spearman’s $ ho$ of 0.403.
  • Loss Function Impact: Switching to the joint MSLE + MAPE loss led to substantial improvements in the MAPE metric, confirming that optimizing for percentage error is more effective for this domain.
  • The Multilingual Paradox: Despite having four times more data, the multilingual model performed slightly worse on English petitions than the monolingual version. This suggests that cultural nuances and varying petitioning behaviors across regions make a universal global model difficult to achieve.

Conclusion and Future Work

This research proves that the visual "vibe" of a petition is a crucial complement to its written plea. Future work will focus on filtering out default images to isolate the impact of original photography and exploring more sophisticated cross-lingual transfer techniques, such as XLM-R, to better account for cultural differences in digital activism.

Summary updated

BibTeX
@inproceedings{kitayama2020popularity,
  author = {Kotaro Kitayama and Shivashankar Subramanian and Timothy Baldwin},
  title = {{Popularity Prediction of Online Petitions using a Multimodal DeepRegression Model}},
  booktitle = {ALTA},
  year = {2020},
}
Figure