A label knowledge graph powered multi-task framework for crowdsourcing and mobile crowd sensing tasks

Crowdsourcing and Mobile Crowd Sensing (MCS) platforms have revolutionized data collection, harnessing the collective intelligence of crowdsourced sensing. Accurately classifying and extracting core information such as heterogeneous sensors in MCS tasks plays a key role in the platform execution efficiency. However, existing methods struggle with extracting pivotal information from task descriptions that are open-domain, implicitly expressed, and linguistically diverse, ultimately hindering the efficiency of task assignment and execution. To overcome these challenges, we propose LKG-MF, a Label Knowledge Graph-powered Multi-task Framework, to achieve better core information mining performance in crowdsourcing and mobile crowd sensing tasks. Specifically, we first construct an MCS task dataset comprising over 10,000 real tasks from 7 platforms. Then we devise a label knowledge graph to capture heterogeneous semantics and relationships among labels and enhance label representation. Further, we present a multi-granularity feature extraction network to capture precise task-specific features. To optimize performance across disparate tasks, we incorporate a task- adaptive loss function that adeptly balances their optimization rates. Experimental results show that LKG-MF outperforms baselines average by 2.3%, significantly improving multi-task classification accuracy. Notably, when we integrate the LKG-MF model into MCS platforms, the task assignment efficiency is improved by 38.6% and the task completion time is reduced by 45.1%, which demonstrates the practical impact and effectiveness of our model in improving the performance of MCS platforms.