A Crowdsourced Learning Framework to Optimize Cross-Event QoS in AI-powered Social Sensing

Social sensing has become a critical source to obtain timely observations of emergent crisis events at an unprecedented scale by exploring the social media data contributed by common citizens. This paper focuses on an AI-powered crisis situation awareness (ACSA) application in social sensing that aims to obtain accurate situation awareness of emergent crisis events by leveraging advanced AI techniques and social sensing data. In particular, we study a cross-event quality-of-service (C-QoS) problem in ACSA applications where the goal is to address the limitation of current AI models that are often optimized only for a single crisis event and lack the generality to provide desirable QoS across different events. This paper explores the integration of AI and crowdsourced human intelligence as a solution to address the limitation in tackling the problem of C-QoS. However, two critical challenges exist: 1) it is challenging to optimize the ACSA model's C-QoS without sacrificing its specificity on each studied crisis event; 2) it is non-trivial to integrate the diversified AI and human intelligence to optimize C-QoS in ACSA applications. To combat these challenges, we introduce CrossGeneral, a subjective logic-driven human-AI collective learning framework that jointly leverages the specificity of AI and the generality of human intelligence to provide robust and accurate ACSA performance across different types of crisis events. Through evaluations performed on two real-world ACSA applications, CrossGeneral exhibits superior performance compared to state-of-the-art baselines by substantially enhancing the C-QoS of ACSA models.