Bridging Psychology and AI: A Full-Cycle Framework for Monitoring Destructive Impacts in Social Networks
Detection and Monitoring of the Destructive Impacts in the Social Networks Using Machine Learning Methods
The paper introduces a comprehensive full-cycle methodology for detecting and monitoring destructive impacts in social networks (specifically VKontakte) targeting young users. It integrates psychological profiling (Ammon’s test) with machine learning classifiers like Multilayer Neural Networks (MNN) and soft voting ensembles to identify behavioral deviations and harmful communities.
TL;DR
Researchers from the St. Petersburg Federal Research Center have developed a holistic methodology to protect young people from digital harm. By merging Ammon’s psychological Ego-functions with Machine Learning, the system detects destructive behavioral shifts in social network profiles and identifies the specific online communities propagating these negative influences.
Background & Motivation: The Digital Vulnerability of Youth
The transition to a digital-first world has exposed young populations to "destructive impacts"—content or interactions that provoke aggression toward others or oneself. The core challenge is twofold:
- Scale: The sheer volume of data makes manual expert monitoring impossible.
- Context: Simple keyword filtering fails to capture the deep-seated psychological shifts that precede aggressive behavior.
The authors argue that the manifestation of a user's psychological state is reflected in their "digital footprint" (posts, images, and community memberships), requiring a bridge between clinical psychology and data science.
Methodology: The Four-Stage Pipeline
The paper proposes a "full-cycle" approach that doesn't just flag content, but monitors the evolution of a user's psychological health.
1. Initial Profiling
The system uses Ammon’s Ego-function test, which measures six personality dimensions (Aggression, Anxiety, Ego-delimitation, etc.). By correlating test results with 100 profile features (textual, numerical, and image-based), they train classifiers to predict a user's psychological state without them taking a physical test.
2. Continuous Monitoring
Profiles are re-evaluated at set intervals. If a user’s "Anxiety" or "Aggression" score exceeds a threshold compared to their baseline, the system flags the profile for manual review and moves to investigate their environment.
3. Community Analysis
Harm doesn't happen in a vacuum. The methodology analyzes the communities (groups) the user belongs to. It uses NLP techniques—including FastText and Skip-gram—combined into a soft voting ensemble to classify whether a community is a source of destructive stimuli.
4. Constructive Intervention
The final (future) stage involves developing "constructive stimuli"—proactive recommendations and counter-narratives to help young people resist identified destructive impacts.
Experimental Insights & Results
The researchers tested various architectures across their datasets:
- For User Profiles: They compared SVM, Linear Regression, and Neural Networks. The Multilayer Neural Network (MNN) outperformed others with 59.94% accuracy. While this may seem low compared to standard CV tasks, in the realm of psychological forecasting from noisy social data, it represents a significant benchmark.
- For Communities: The soft voting ensemble proved most robust, achieving 55.72% accuracy in identifying harmful groups based on post content.
Performance Comparison
| Model | Profile Accuracy | Community Accuracy |
|---|---|---|
| SVM / Basic Classifiers | Lower | ~50% |
| MNN / Soft Voting Ensemble | 59.94% | 55.72% |
Critical Perspective: Why This Matters
The true value of this work is not in achieving a 99% accuracy score—which is likely impossible given the subjectivity of psychology—but in its integrated architecture.
Strengths:
- Psychological Grounding: Unlike "black-box" toxicity filters, this method uses established Ego-function scales (Ammon's), providing a theoretical basis for why a profile is flagged.
- Proactive Monitoring: It identifies the change in a person's digital behavior over time, acting as an early warning system.
Limitations:
- Accuracy Floor: The accuracy (~60%) necessitates a human-in-the-loop. It cannot be used for autonomous censorship or intervention without significant risk of false positives.
- Platform Specificity: The current study is localized to Vkontakte; cross-platform dynamics (TikTok, Instagram) remain unexplored.
Conclusion
This paper represents a vital step toward a "healthy digital environment." By treating social networking not just as a data stream but as a reflection of the human psyche, the authors provide a template for how AI can move from simple content filtering to complex social protection. Future work on "constructive stimuli" will be the litmus test for whether this methodology can move from diagnosis to cure.