Rethinking Posts Through Emotion Awareness: A Multi-Agent Approach to Social Privacy

TL;DR

This research introduces an emotion-aware Multi-Agent System (MAS) that acts as a psychological safety net for social media users. By analyzing both facial expressions (to build a PAD model) and the sentiment of written text, the system predicts when a user is in a vulnerable "temperament state" and intervenes before they share content they might later regret.

Background & Positioning

In the landscape of digital privacy, most tools focus on encryption or access control lists. This paper shifts the focus to behavioral privacy. It addresses the "regret loop" common in social media interactions—specifically among teenagers—positioning itself as a proactive interventionist tool that bridges affective computing and social network security.

Problem & Motivation: The Vulnerability of the "Heat of the Moment"

The authors identify a significant gap in current social platforms: they are "emotionally blind." Research shows that users, particularly adolescents, are more likely to make risky disclosures or aggressive posts when their emotional state is compromised. Current systems allow these posts to go live instantly, leading to long-term social and privacy consequences. The goal here is to introduce a "cool-down" mechanism driven by the user's own emotional data.

Methodology: The Logic of Emotion-Aware Agents

The core of the proposal is a three-layer Multi-Agent System (MAS) that processes data through specialized agents:

1. PAD Calculator: Extracts facial features from user images to map them onto the Pleasure-Arousal-Dominance (PAD) space. This provides a objective metric of the user's current mood.
2. Sentiment Analysis Agent: Performs aspect-based sentiment analysis on the text of a pending post using an Artificial Neural Network (ANN) to determine polarity.
3. User Model Calculator: The "brain" of the system, which synthesizes PAD data and text polarity into a Temperament Model based on Mehrabian’s theories.
4. Advisor: The gatekeeper. If the system detects a "Negative User State + Negative Post Sentiment" collision, it triggers a warning.

Figure 1: The interaction between the User Agent and the Advisor Agent during the decision-making process.

Preliminary Results & Future Work

The system is currently in the prototype stage, moving toward integration into a real-world social network tailored for teenagers. The preliminary architecture demonstrates that a multi-modal approach (Visual + Textual) provides a more holistic view of user intent than text analysis alone.

The authors plan to validate the system through a multi-week study with real teenage users to measure the reduction in "regretful" or high-risk posts.

Critical Insight & Conclusion

The significance of this work lies in its Inductive Bias: it assumes that privacy is not just a technical setting, but a byproduct of emotional state. While the reliance on facial images for PAD modeling raises its own set of privacy questions (which the authors address via the Persistence Layer agent), the integration of psychological temperament into Multi-Agent Systems marks a sophisticated step forward in human-centric AI.

Future Outlook: For this to be viable at scale, the system must address the "privacy paradox" of monitoring a user's face to protect their posts. However, as an educational tool for secondary education, it offers a powerful framework for teaching digital literacy through algorithmic reflection.