TLDR: This research introduces a new AI system that uses social media data to identify and classify estate-related events and topics. The system employs a hierarchical classification framework, first filtering relevant posts and then categorizing them into actionable topics like infrastructure, parking, and noise. It also includes a geolocation module to infer posting locations for posts without explicit geotags. This integrated approach provides timely, data-driven insights for urban management and operational response, helping to address information overload on social media platforms.
Social media platforms like Twitter/X and Facebook have become an integral part of our daily lives, serving as dynamic sources of local news and personal experiences. With urban populations growing, these platforms are increasingly valuable for identifying issues related to estates and urban management. However, the sheer volume and speed of information on social media present a significant challenge: information overload, making it difficult to find relevant content efficiently.
Addressing this challenge, a new system has been developed to automatically detect and classify estate-related events from social media content. This innovative system aims to bridge a critical gap in social media analysis, specifically focusing on issues such as facility breakdowns, noise complaints, or parking violations, whether from historical data or real-time streams. This initiative is part of a larger effort called Estate-IQ, which seeks to automate estate operations and maintenance through AI-driven event detection and decision support.
How the System Works
The proposed system operates through a sophisticated, integrated approach, designed to first filter relevant posts and then categorize them into actionable estate-related topics. It comprises four main components:
- Data Repository/Stream: This component handles the incoming social media posts, whether from a stored collection or a live feed. The system can process these posts in real-time or in batches, supporting both past and current analyses.
- Estate-related Post Classification: To identify content relevant to estate management, the system uses a powerful language model called BERT (Bidirectional Encoder Representations from Transformers). This model is trained to determine if a post contains estate-related information (e.g., facility faults, noise complaints) or not. Posts identified as relevant are then passed on for further categorization.
- Estate Topic Classification: Once a post is confirmed as estate-related, this component further classifies it into specific topics. Using another fine-tuned BERT model, the system categorizes posts into four main areas: Infrastructure, Parking, Noise, and a general ‘Others’ category for less frequent issues. This helps in pinpointing the exact nature of the estate-related problem.
- Social Post Geolocation: For posts that don’t explicitly include location information (geotags), an additional module is used to infer where the post likely originated. This transformer-based geolocation framework, called transTagger, uses both text and other cues to estimate the posting location, even down to a neighborhood level. This feature is crucial for spatially contextualizing posts, aiding in real-time urban monitoring and mapping of events.
Performance and Impact
Preliminary evaluations of the system’s core components have shown promising results. For detecting estate-related posts, the BERT model achieved high accuracy and F1-scores, outperforming other methods. In classifying estate topics, BERT also performed strongly for Infrastructure, Parking, and Noise categories. While the ‘Others’ category had lower performance, it represents a small portion of the data and doesn’t significantly impact the overall effectiveness.
The geolocation module, transTagger, demonstrated reasonable accuracy in inferring locations, especially considering the complexity of identifying specific points of interest. To protect user privacy, the system infers locations at a broader neighborhood level rather than precise points.
Also Read:
- Mapping Research Fields with AI: The Sci-OG Methodology
- AI Conferences at a Crossroads: A New Model for Sustainable Research Exchange
Conclusion
This social data-driven platform represents a significant step forward in leveraging social media for urban management. By automatically identifying and categorizing estate-related posts, the system helps overcome information overload, providing timely and actionable insights for urban operations and decision-making in densely populated areas. Future work will explore privacy implications and potential biases in both classification and geolocation tasks.
For more in-depth information, you can read the full research paper: A Social Data-Driven System for Identifying Estate-related Events and Topics.
