Abstract
Social media platforms often trap users in repetitive content loops, a phenomenon known as "cognitive-behavioral fixation." This paper introduces a novel computational framework to quantify this state by analyzing multimodal data (text and video). By combining hierarchical topic extraction via Vision-Language Models with metrics for diversity, dominance, and recurrence, the system identifies when browsing habits transition from healthy exploration to obsessive obsession. Validated on a large-scale dataset of 1.1 million records, the framework achieves 85.7% accuracy in identifying fixated behavior.
Introduction
Modern social media algorithms are optimized for engagement, frequently creating self-reinforcing feedback loops. While concepts like "echo chambers" are well-documented, there is a lack of unified methods to measure the broader psychological state of cognitive-behavioral fixation—an obsessive preoccupation that limits flexible thinking. Traditional psychological assessments rely on manual surveys, which are unscalable for real-time digital interactions. This research bridges the gap by proposing an automated, multimodal framework to detect fixation patterns in massive social media datasets.
Methodology: Multimodal Hierarchical Topic Extraction
The proposed system operates in two primary stages: topic extraction and fixation quantification.
1. Multimodal Topic Extraction
Unlike traditional Latent Dirichlet Allocation (LDA), which struggles with the brevity and noise of social media, this framework utilizes a Vision-Language Model (MiniCPM-V). The model processes both text descriptions and video frames to generate fine-grained "first-level" topic tags. These tags are then embedded using Sentence-BERT and clustered via MiniBatch K-Means into broader "second-level" thematic categories, such as "Space Technology" or "Culinary Delights."
2. Fixation Quantification
The system calculates a "Combined Fixation Score" based on three behavioral metrics:
- Diversity (Shannon Entropy): Quantifies the breadth of a user's interests. Lower entropy indicates a narrower focus.
- Dominance (Herfindahl-Hirschman Index): Measures the concentration of attention on a single topic, borrowed from economic market share analysis.
- Recurrence (Burstiness): Captures the temporal frequency of engagement, identifying users who repeatedly return to the same topic in short windows.
Data and Experimental Insights
The framework was evaluated using the X User Browsing Dataset (XUB), a curated collection of 1.1 million records from 163 anonymized users. This dataset is unique for its inclusion of temporal sequences, allowing researchers to track the development of fixation over time.
Key findings include:
- Multimodal Efficacy: Integrating video frames with text improved topic coherence to 75.6%, as visual data helps disambiguate vague or sarcastic social media text.
- Fixation as a Spectrum: Most users maintain high diversity (peaking at 0.87 entropy). Fixation is characterized by a sharp drop in diversity and a corresponding rise in dominance.
Results
The framework demonstrated high performance, achieving an accuracy of 85.7% and an F1-score of 0.667 against human-annotated ground truth. Analysis of the XUB dataset revealed that approximately 8.59% of users exhibited fixated behavior, with common domains including sports, gaming, and cooking. Notably, "fixation streaks" were observed to last up to 24 consecutive days, highlighting the persistence of these behavioral loops.
Conclusion and Future Work
This research provides a scalable tool for identifying obsessive digital behavior. However, limitations remain, such as the exclusion of audio data and interaction metrics like "dwell time." Future iterations could refine the scoring system by weighting metrics based on specific mental health outcomes and incorporating causal inference to distinguish between user predisposition and algorithmic influence.
Summary updated