[Paper Review] Independent Component Analysis for Trustworthy Cyberspace during High Impact Events: An Application to Covid-19

Zois Boukouvalas, Christine Mallinson+6 more|arXiv (Cornell University)|Jun 30, 2020

Anomaly Detection Techniques and Applications27 references4 citations

TL;DR

This paper proposes a data-driven independent component analysis (ICA) framework for joint detection of misinformation and knowledge discovery in social media during high-impact events, applying it to a novel, expert-labeled COVID-19 Twitter dataset. The method extracts interpretable linguistic features from tweet word counts, achieving effective classification of reliable vs. unreliable content while revealing semantic patterns linked to misinformation characteristics such as conspiracy theories and emotional language.

ABSTRACT

Social media has become an important communication channel during high impact events, such as the COVID-19 pandemic. As misinformation in social media can rapidly spread, creating social unrest, curtailing the spread of misinformation during such events is a significant data challenge. While recent solutions that are based on machine learning have shown promise for the detection of misinformation, most widely used methods include approaches that rely on either handcrafted features that cannot be optimal for all scenarios, or those that are based on deep learning where the interpretation of the prediction results is not directly accessible. In this work, we propose a data-driven solution that is based on the ICA model, such that knowledge discovery and detection of misinformation are achieved jointly. To demonstrate the effectiveness of our method and compare its performance with deep learning methods, we developed a labeled COVID-19 Twitter dataset based on socio-linguistic criteria.

Motivation & Objective

• To address the challenge of detecting misinformation in social media during high-impact events like the COVID-19 pandemic.
• To overcome limitations of handcrafted features and black-box deep learning models by developing a data-driven, interpretable approach.
• To jointly achieve misinformation detection and knowledge discovery through latent variable modeling.
• To create and release a novel, expert-annotated COVID-19 Twitter dataset labeled using socio-linguistic criteria for reproducible research.

Proposed method

• Formulate tweet classification as a blind source separation problem using the ICA model X = AS, where X is the word count matrix, A is the mixing matrix, and S is the source signal matrix.
• Apply dimensionality reduction via PCA to reduce the feature space when d > N, ensuring a well-posed ICA problem.
• Estimate the demixing matrix W to recover maximally independent source components y(v) = Wx(v), representing latent linguistic features.
• Use the resulting independent components as low-dimensional, interpretable features for training a binary classifier distinguishing reliable and unreliable tweets.
• Perform knowledge discovery by analyzing the top-weighted words in each estimated source component to interpret semantic content.
• Construct a labeled dataset of 1,000 COVID-19-related tweets using subject-matter experts and socio-linguistic criteria, including features like emotional language, conspiracy theories, and false health advice.

Experimental results

Research questions

• RQ1 Can ICA-based feature extraction effectively distinguish between reliable and unreliable tweets during the COVID-19 pandemic?
• RQ2 How do the interpretable linguistic components derived from ICA relate to known characteristics of misinformation, such as emotional language or conspiracy theories?
• RQ3 To what extent can a data-driven, non-deep learning method achieve competitive performance compared to state-of-the-art deep learning models in misinformation detection?
• RQ4 How can socio-linguistic criteria be systematically applied to create reliable, expert-annotated datasets for misinformation research?
• RQ5 What insights into the propagation of misinformation can be gained by analyzing the semantic content of independent components?

Key findings

• The ICA-based method achieved effective classification of reliable and unreliable tweets using only word count features, demonstrating strong performance without deep learning architectures.
• The extracted independent components revealed distinct semantic clusters: Feature 1 related to Olympic cancellation, Feature 5 to New York case counts and emergency declarations, and others to specific misinformation types.
• Features associated with unreliable content were strongly aligned with known linguistic markers of misinformation, such as political bias (Feature 5), emotional language (Feature 4), and false health advice (Feature 3).
• The method successfully identified and localized key misinformation narratives, such as the 'Pirbright Institute' conspiracy theory and 'Bill Gates patent' rumors, through high-weighted terms in the mixing matrix.
• The labeled dataset, constructed using socio-linguistic criteria, provides a reliable benchmark for future research and enables reproducible evaluation of misinformation detection systems.
• The interpretability of ICA components allows analysts to trace how specific linguistic patterns correlate with misinformation, supporting explainable AI in crisis communication.