Boosting urban prediction tasks with domain-sharing knowledge via meta-learning

Urban prediction tasks refer to predicting urban indicators ( e.g. , traffic, temperature, etc.) using urban big data, which is crucial for understanding the urban patterns, and further benefits the urban public administration. An empirical study indicates that there are correlated patterns among urban prediction tasks from various domains, which suggests the existence of domain -sharing knowledge. Aggregating such domain -sharing knowledge would significantly benefit urban prediction tasks. However, as a widely used learning paradigm for knowledge aggregation, existing meta -learning methods, especially gradient -based methods, can only work for singledomain tasks. To solve the problem, we propose Cross -Domain Meta -Learning (CDML), a flexible framework for aggregating domain -sharing knowledge from cross -domain urban prediction tasks. Specifically, the core architecture of CDML is the model fusion block that includes (1) meta -model, shared by cross -domain tasks for capturing domain -sharing knowledge; (2) domain -specific model, shared only by the same -domain tasks for preserving domain -specific knowledge; and (3) knowledge fusion unit, for combining both the domainsharing/specific knowledge for good generalization. Moreover, we develop asynchronous meta -training and adaption strategy strategies to further guarantee cross -domain generalization. The extensive experimental results validate the effectiveness of the proposed framework with the superior ability of boosting existing urban prediction models, quick adaption, and the potential for simplifying models.