Invisible Walls in Cities: Designing LLM Agent to Predict Urban Segregation Experience with Social Media Content

Technical Summary: Invisible Walls in Cities: Designing LLM Agent to Predict Urban Segregation Experience with Social Media Content

Problem Formulation
The paper addresses the challenge of predicting "experienced segregation"—the dynamic barriers individuals face in daily urban movements—using unstructured social media data. While historical segregation research focused on residential patterns, modern urban spaces present "invisible walls" driven by cultural, social, and economic factors that limit interaction between demographic groups. The core problem is formulated as a Point of Interest (POI) segregation prediction task. Given a set of POIs, their associated social media content (reviews, images, attributes), and local demographic data, the objective is to learn a function that minimizes the discrepancy between the predicted segregation index and the ground truth. The ground truth is quantified as the deviation between the racial composition of a POI's visitors and the racial composition of the surrounding residential population.

Methodology
The authors propose a two-stage framework: a Reflective LLM Coder for qualitative insight extraction and a REasoning-and-EMbedding (RE'EM) framework for quantitative prediction.

1. Reflective LLM Coder:
   To overcome the ambiguity and volume of raw social media data, the authors design an agentic workflow to generate a structured codebook.

   • Reflective Attributor: This module employs a two-step Chain-of-Thought (CoT) process. First, it performs multi-source insight generation by analyzing POI names, reviews, and images to identify factors influencing appeal to specific racial/ethnic groups. Second, it applies reflection-guided insight refinement, where the LLM compares its initial insights against real-world visitor demographic data (the "expectation gap"). This abductive reasoning step filters insights to retain only those consistent with observed visitation patterns.
   • Code Summarizer: This module aggregates the refined insights into a structured, orthogonal codebook through four steps: extraction, semantic combination, quality-based removal, and explanation refinement. The resulting codebook identifies nine key dimensions (e.g., Cultural Resonance, Price Sensitivity, Accessibility, Community Engagement) that act as "bricks" forming invisible walls.

2. REasoning-and-EMbedding (RE'EM) Framework:
   This framework integrates three complementary channels to predict segregation scores:

   • Reasoning Channel: Guided by the generated codebook, an LLM simulates the perspective of five distinct racial/ethnic groups to rate each POI on the nine codebook dimensions (0–10 scale). These structured ratings are vectorized via a Reasoning Adapter (MLP).
   • Embedding Channel: A fine-tuned text embedding model (GTE-base) processes the raw review corpus to extract holistic semantic representations, which are then passed through an Embedding Adapter.
   • Population Channel: This channel incorporates static demographic context by computing the weighted average racial composition of Census Block Groups within a 0.5km radius of the POI, transformed via a Population Adapter.
   • Neighbor-Aware Multi-View Predictor: The final predictor utilizes Graph Attention Networks (GAT) to aggregate features from the three channels. It identifies nearest neighbors in three distinct spaces (reasoning similarity, embedding similarity, and geographic proximity) and fuses the POI's own representations with its neighbors' aggregated representations to output the final segregation prediction.

Key Contributions

• Novel Application: This is the first work to utilize social media content for predicting POI-level experienced segregation.
• Reflective LLM Coder: The authors design a novel agent that automates the mining of online reviews, using real-world feedback to refine insights and produce a generalizable codebook for segregation factors.
• RE'EM Framework: A hybrid architecture that combines the reasoning capabilities of LLMs (via structured ratings) with the representational power of embedding models and demographic data, fused through a neighbor-aware mechanism.
• Comprehensive Validation: The approach is validated through both quantitative experiments on real-world data and a qualitative user study.

Results

• Quantitative Performance: Experiments on four US cities (Philadelphia, Tucson, Tampa, New Orleans) demonstrate that the RE'EM framework significantly outperforms baseline models relying solely on local racial composition. The proposed method achieves a 22.79% increase in $R^2$ and a 9.33% reduction in Mean Squared Error (MSE). The derived codebook was shown to be generalizable across different cities, consistently improving prediction accuracy.
• User Study: In a study with 75 researchers, participants provided with LLM-generated, codebook-guided summaries showed significantly improved prediction accuracy compared to those using vanilla LLM outputs. Furthermore, 80% of participants preferred the codebook-guided summaries, citing them as providing greater cognitive gain in perceiving the social inclusiveness of POIs.

Significance
The paper claims to mark an important step toward understanding implicit social barriers and inequalities in urban environments. By demonstrating the potential of Web technology and Large Language Models to extract operationable insights from unstructured data, the study offers a pathway for promoting social inclusiveness. The work suggests that automated analysis of digital traces can inform individuals about potentially uncomfortable situations and support policymakers in fostering more equitable cities, moving beyond retrospective measurements to predictive capabilities.