QuickStop: A Markov Optimal Stopping Approach for Rapid Misinformation Detection

Description

Background

The proliferation of misinformation (colloquially known as "fake news'') on online social networks has become one of the greatest threats to our national security, to public trust in news media, and to the ecosystem of online social platforms. Third-party fact-checking methods are often very effective for analysis of a specific news article but are not scalable solutions. The crowdsourcing approach is more scalable, but trustworthiness can be questionable because anyone submit a report. In light of these challenges, machine-learning and data-mining approaches have emerged as potential tools to handle misinformation detection in a systematic and comprehensive manner. It has been shown that (1) the features extracted from the content of a news article, (2) the features of the users who spread the news, and (3) the connections of these users can all be effectively utilized for misinformation detection. Capitalizing on these advancements can therefore result in a highly scalable solution that can process a vast number of news articles in a short period of time.

Invention Description

Researchers at Arizona State University have developed QuickStop, a misinformation detection algorithm framed as a Markov optimal stopping problem with an asymmetric cost function towards misinformation. Using a threshold-based stopping rule from martingale theory, tests with real-world data demonstrated that QuickStop outperformed existing algorithms even though the latter used 10 times (and sometimes 50 times) more observations as well as more features. Furthermore, numerical evaluation with synthetic data showed that the algorithm is robust to edge classification errors.

Practically speaking, consider an online social network platform monitoring the spread of some information: When a user “retweets” or posts the information, QuickStop uses this weak signal to update the current system state while keeping complexity independent of how many observations have been collected. Then by comparing the state with several thresholds calculated offline, the algorithm decides whether to keep collecting observations or declare the type of information.

Potential Applications

• Social media

• Fact-checking

Benefits and Advantages

• Combines data-driven and model-driven methods for real-time misinformation detection

• Optimizes the number of required observations for improved real-time operation

• Emphasizes scalability, accuracy, and speed

Related Publication

Case ID:
M19-292P^
Published:
06-25-2020
Last Updated:
06-25-2020

For More Information, Contact