[Paper Review] Human as Real-Time Sensors of Social and Physical Events: A Case Study of Twitter and Sports Games

Siqi Zhao, Lin Zhong+2 more|arXiv (Cornell University)|Jun 21, 2011

Video Analysis and Summarization17 references91 citations

TL;DR

This paper proposes using Twitter as a real-time sensor network for detecting social and physical events, demonstrating its feasibility through a case study of NFL games. By leveraging streaming data and event recognition algorithms, the system detects game events within 40 seconds with up to 90% accuracy, enabling applications in live media guidance and targeted advertising.

ABSTRACT

In this work, we study how Twitter can be used as a sensor to detect frequent and diverse social and physical events in real-time. We devise efficient data collection and event recognition solutions that work despite various limits on free access to Twitter data. We describe a web service implementation of our solution and report our experience with the 2010-2011 US National Football League (NFL) games. The service was able to recognize NFL game events within 40 seconds and with accuracy up to 90%. This capability will be very useful for not only real-time electronic program guide for live broadcast programs but also refined auction of advertisement slots. More importantly, it demonstrates for the first time the feasibility of using Twitter for real-time social and physical event detection for ubiquitous computing.

Motivation & Objective

• To explore the feasibility of using Twitter as a real-time sensor for social and physical events.
• To address challenges in real-time data access and event detection due to rate limits and data availability constraints.
• To develop an efficient, scalable system for detecting diverse events using Twitter streams.
• To evaluate the system’s performance in a real-world context using live NFL games.

Proposed method

• The authors designed a web service that ingests real-time Twitter streams using the public API, despite rate limitations.
• They implemented a lightweight event recognition pipeline that identifies event markers through temporal and linguistic patterns in tweets.
• The system uses a combination of keyword matching, temporal clustering, and sentiment analysis to detect game-related events such as touchdowns or timeouts.
• A sliding window mechanism processes incoming tweets to detect event spikes and reduce noise.
• The solution is optimized for low-latency processing, achieving detection within 40 seconds of event occurrence.
• The system was deployed and evaluated during the 2010–2011 NFL season, using actual game data for validation.

Experimental results

Research questions

• RQ1 Can Twitter be effectively used as a real-time sensor for detecting physical and social events?
• RQ2 How accurately and quickly can event detection be achieved under real-world data access constraints?
• RQ3 What techniques enable reliable event recognition in noisy, high-velocity social media streams?
• RQ4 Can such a system support practical applications like live media guidance and targeted advertising?
• RQ5 What are the performance trade-offs between detection speed, accuracy, and system scalability?

Key findings

• The system successfully detected NFL game events within an average of 40 seconds of occurrence.
• Event detection accuracy reached up to 90% in the evaluation phase.
• The system remained robust despite rate limits and incomplete data access from Twitter’s public API.
• Temporal clustering and linguistic pattern analysis significantly improved detection precision over simple keyword matching.
• The solution demonstrated scalability and real-time performance suitable for production deployment.
• The study provides the first empirical evidence of Twitter’s feasibility as a real-time sensor for ubiquitous computing applications.