EventGo!: Decoding Urban Dynamics through Social Media Information Extraction

TL;DR

EventGo! is a system designed to transform the chaotic stream of Facebook fan pages into a structured, searchable database of local events. By leveraging Locality Sensitive Hashing (LSH) and a Double-Tier Automatic Labeling strategy, the researchers successfully extracted event names and venues from over 230,000 pages, providing a vital tool for small business discovery and urban exploration.

Problem & Motivation: The "Fragmented Event" Dilemma

While high-end concerts and major festivals are easily found on ticketing websites, the heartbeat of a city often lies in smaller, "long-tail" events—promotions at local cafes, community workshops, or pop-up markets.

Currently, these events are predominantly shared as unstructured posts on social networking sites like Facebook. For users, these are nearly impossible to search systematically. For researchers, the challenge is two-fold:

1. Data Volume: Processing hundreds of thousands of active fan pages.
2. Labeling Bottleneck: Traditional NLP models require massive amounts of labeled data, which is expensive and slow to produce manually for niche event categories.

Methodology: Scalable Extraction via Distant Supervision

The core innovation of EventGo! lies in its ability to generate high-quality training data without manual labor.

1. Web Scraping & Data Cleaning

The pipeline begins by crawling 230K Facebook fan pages. Given the diversity of posting styles, the system focuses on identifying text blocks that likely contain "Event" identifiers.

2. LSH and Seed-List Distant Supervision

To overcome the "cold start" problem in labeling, the authors use Distant Supervision. They utilize existing seed lists of known venues and event types to label the raw text automatically. To speed this up and handle near-duplicate posts (common in social media marketing), they employ Locality Sensitive Hashing (LSH).

3. Double-Tier Automatic Labeling

To ensure the training data isn't just "large" but also "accurate," the system uses a two-stage filtering process:

• Tier 1: Initial labeling based on string matching and seed lists.
• Tier 2: A refinement stage that uses the context of the posts to prune false positives, ensuring the Recognition (NER) model learns from high-precision examples.

Note: The system integrates scraping, cleaning, and the labeling pipeline to feed the final search index.

Experiments & Results

The researchers disclosed statistics showing a significant shift in the advertising market, where small-to-medium enterprises (SMEs) have moved almost entirely to social media "posts" rather than "official events."

By applying their Double-Tier Labeling, they achieved a dramatic reduction in training data preparation time. More importantly, the system proved robust enough to distinguish between "Event Venues" and simple "Location Tags," which is a common failure point for standard geolocation tools.

Critical Analysis & Conclusion

Takeaway: EventGo! proves that for "hidden" data in social media, the bottleneck isn't just the model architecture—it's the data engineering. By using smart hashing and distant supervision, they built a SOTA service for real-world urban exploration.

Limitations: Since the system relies on Facebook Fan Pages, it is subject to the platform's API limitations and data privacy changes. Furthermore, the reliance on "Seed Lists" for distant supervision means the system might struggle with entirely new, "vibe-based" event categories that don't match traditional keywords.

Future Outlook: This framework could easily be extended to other platforms like Instagram or TikTok, where visual cues could supplement the text-based event extraction, creating an even more immersive "City Dynamics" map.