Predicting Stances from Social Media Posts using Factorization Machines

Akira Sasaki, Kazuaki Hanawa, Naoaki Okazaki, Kentaro InuiCOLING 2018 · OpenReview

Abstract

Social media provide platforms to express, discuss, and shape opinions about events and issues in the real world. An important step to analyze the discussions on social media and to assist in healthy decision-making is stance detection. This paper presents an approach to detect the stance of a user toward a topic based on their stances toward other topics and the social media posts of the user. We apply factorization machines, a widely used method in item recommendation, to model user preferences toward topics from the social media data. The experimental results demonstrate that users’ posts are useful to model topic preferences and therefore predict stances of silent users.

Introduction

Stance detection—the task of determining whether a user is in favor of or against a specific topic—is a critical component of social media analysis. However, traditional text classification approaches struggle with two primary challenges. First, there is an explicit statement gap: users often express opinions that imply a stance without explicitly mentioning the topic. For example, a user might advocate for protecting domestic agriculture, which indirectly signals a disagreement with a trade agreement like the Trans-Pacific Partnership (TPP).

Stance detection challenge: Implicit stances vs. explicit topics.

Second, the "silent majority" problem persists; a vast number of users are "lurkers" who rarely post about political or social issues. In a dataset of over 444,000 users, approximately 70% never explicitly stated a stance on any of the 1,000+ topics tracked. While Matrix Factorization can predict missing stances for active users, it fails for those with no prior known stances.

Methodology

The authors propose treating stance detection as a recommendation problem, utilizing Factorization Machines (FMs). FMs are particularly effective at modeling interactions between diverse feature sets, making them ideal for integrating structured stance data with unstructured textual content.

Framework for stance prediction using Factorization Machines (FMs).

The model's input for a given (User, Topic) pair includes:

  • User and Topic Identifiers: To capture individual user biases and general topic trends.
  • Known Stances: A vector representing the user's stances on other topics.
  • User Posts: Textual features (n-grams and linguistic dependencies) extracted from the user's general social media history.

To establish ground truth, the researchers used 200 high-precision linguistic patterns to automatically extract explicit stances from a corpus of 1.76 billion Japanese tweets.

Key Results and Insights

The study reveals that a user's general "stream of consciousness" serves as a vital latent context for their beliefs. Even without mentioning a specific policy, the vocabulary and phrasing in everyday posts correlate with underlying worldviews.

  • Predicting the Silent: For users with zero explicit stances, the FM model utilizing "Posts" significantly outperformed the majority baseline.
  • The 500-Tweet Threshold: Prediction accuracy for silent users was found to plateau after analyzing approximately 500 general tweets, suggesting that a relatively small digital footprint is sufficient to characterize a user's likely stance.
  • Performance Gains: In the task of completing missing stances, the full model (Topic + User + Other Stances + Posts) achieved 66.45% accuracy, surpassing both the majority baseline (63.67%) and standard Matrix Factorization (61.12%).

Discussion and Future Work

While the FM model excels at stance prediction, it occasionally relies on textual "shortcuts" rather than deep logical connections between topics. Future improvements could involve using distributed representations (such as sentence embeddings) to address feature sparsity. Additionally, incorporating social signals—such as follow/follower relationships and retweets—could further enhance the model's predictive power by capturing the social context of opinion formation. Ultimately, this work highlights that including general post data is essential for representative social media analysis, as it allows researchers to account for the viewpoints of the silent majority.

Summary updated

Abstract

Social media provide platforms to express, discuss, and shape opinions about events and issues in the real world. An important step to analyze the discussions on social media and to assist in healthy decision-making is stance detection. This paper presents an approach to detect the stance of a user toward a topic based on their stances toward other topics and the social media posts of the user. We apply factorization machines, a widely used method in item recommendation, to model user preferences toward topics from the social media data. The experimental results demonstrate that users’ posts are useful to model topic preferences and therefore predict stances of silent users.
BibTeX
@inproceedings{sasaki2018predicting,
  author = {Akira Sasaki and Kazuaki Hanawa and Naoaki Okazaki and Kentaro Inui},
  title = {{Predicting Stances from Social Media Posts using Factorization Machines}},
  booktitle = {International Conference on Computational Linguistics},
  year = {2018},
  url = {https://openreview.net/forum?id=7rP4IRPjCgz},
}
Figure