Event Detection in Social Networks Using Polarization Dynamics

Technical Summary: Event Detection in Social Networks Using Polarization Dynamics

Problem Statement
Social networks serve as critical spaces for public deliberation, particularly during politically sensitive moments where user interactions can reveal societal tensions and polarization. While existing event detection methods in social media primarily rely on quantitative spikes in post volume or shifts in sentiment polarity, these approaches may overlook events characterized by deep disagreement or structural fragmentation that do not necessarily generate high traffic volumes or extreme emotional peaks. This paper addresses the challenge of detecting significant public events by leveraging polarization dynamics as a complementary signal, specifically asking: How can a model for event detection be built based on a polarization metric applied to social media data?

Methodology
The study employs a retrospective, statistical approach using data from X (formerly Twitter) regarding the public debate on the 2024 Colombian pension reform (June–July 2024). The methodology follows a structured pipeline:

1. Data Collection and Preprocessing: A dataset of 9,997 tweets was collected via the X API, filtered for relevance to the pension reform, and cleaned of duplicates and non-Spanish content. Unlike traditional workflows, heavy normalization (e.g., lowercasing, tokenization) was avoided to preserve stylistic cues beneficial for stance estimation.
2. Stance Classification: A Large Language Model (GPT-4o) was utilized to classify the stance of each post on a five-point Likert scale (Strongly Disagree to Strongly Agree). Contextual information, including original tweets for replies, was provided to ensure consistent inference.
3. Signal Derivation: From the stance classifications, three daily time series were constructed:
   • Tweet Volume: The raw count of posts.
   • Sentiment Proportions: Derived directly from stance polarity (disagreement as negative, agreement as positive).
   • Polarization: Measured using the Minimum Effort to Consensus (MEC) index. This metric calculates the minimal weighted effort required to shift an opinion distribution toward a consensus point, capturing the cost of fragmentation.
4. Event Detection Models: Three models were compared using anomaly detection on daily changes:
   • Volume-based Model: Detects bursts in posting activity using a sliding window and threshold.
   • Sentiment-based Baseline: Detects abrupt shifts in the proportion of positive/negative posts (adapted from [12]).
   • Proposed Polarization-based Model: Identifies significant shifts in the daily change of the MEC index. This model applies an Interquartile Range (IQR) criterion ($\Delta_t > Q_3 + \lambda \cdot IQR$ or $\Delta_t < Q_1 - \lambda \cdot IQR$) to detect anomalous deviations without assuming strict normality.
5. Evaluation: The sentiment-based model's detections served as a proxy set of reference event labels. Performance was evaluated using Precision, Recall, and F1-score against this baseline.

Key Results
The comparative evaluation yielded the following performance metrics:

• Polarization-based Model: Achieved an F1-score of 0.75 (Precision: 0.75, Recall: 0.75).
• Tweet Volume-based Model: Achieved an F1-score of 0.53 (Precision: 0.71, Recall: 0.42).
• Sentiment-based Baseline: Served as the reference standard (12 detected events after parameter tuning).

Qualitative analysis revealed that the polarization model successfully identified events that were missed by volume and sentiment signals. For instance, on June 27, a polarization spike was detected corresponding to news about private pension funds drafting secret amendments, despite no surge in volume or sentiment. Conversely, on July 16 (the day the reform was signed), a volume spike occurred without a corresponding polarization shift. The results indicate that polarization dynamics capture episodes of increased disagreement and social tension that are not necessarily accompanied by high activity or sentiment peaks.

Significance and Claims
The paper claims that polarization dynamics, operationalized through the MEC index, provide a valuable and complementary signal for monitoring contentious public discourse. The authors argue that while volume and sentiment models effectively capture attention bursts and emotional peaks, they may miss events where the primary signature is a reconfiguration of agreement and disagreement. By bridging event detection and polarization analysis, this approach offers a unified framework to identify socially meaningful events driven by structural divergence rather than just activity levels. The study concludes that relying on a single signal (volume or sentiment) may provide an incomplete view of social dynamics, and that integrating polarization metrics can offer early indications of emerging tensions or shifts in collective positioning in digital environments.