More about HKUST
Selective Multi-Source, Non-IID Transfer Learning for Urban Spatio-temporal Forecasting
PhD Thesis Proposal Defence Title: "Selective Multi-Source, Non-IID Transfer Learning for Urban Spatio-temporal Forecasting" by Mr. Yilun JIN Abstract: Spatio-temporal Forecasting (STF) is a fundamental problem in urban computing, supporting a wide range of downstream tasks like transportation, energy, meteorology, public security, etc. In recent years, deep learning methods have achieved tremendous success in urban STF. However, the constant urbanization process creates needs for new cities, new urban services, and expansion of existing services. In these scenarios, the problem of data scarcity arises, which negatively impacts the accuracy of urban STF and the corresponding downstream applications. Transfer learning is an effective method to address the data scarcity issue in fields like computer vision, natural language processing, and recommendation. However, the spatial and temporal heterogeneity in urban spatio-temporal data poses unique challenges to applying transfer learning for urban STF. Specifically, due to such heterogeneity, the source city (or cities) contains multiple non-IID latent domains, resulting in multi-source, non-IID transfer learning problems. Consequently, it is necessary that some latent source domains are helpful to the target city, while others may be irrelevant or harmful, which compromises the performance of transfer learning for urban STF. To address this problem, in this thesis proposal, we focus on the problem of selective multi-source, non-IID transfer learning for urban STF, whose goal is to identify and select helpful latent domains from the source city and transfer the knowledge to the target city. We first define the problem of selective transfer learning for urban STF and identify three practical sub-problems: selective transfer learning for few-shot grid-like urban STF, few-shot graph-like urban STF, and zero-shot urban STF. We then present two novel selective transfer learning methods, CrossTReS and TransGTR, for few-shot grid-like and graph-like urban STF, respectively. CrossTReS selects latent source domains by adaptively re-weighting source spatial units. By assigning helpful source spatial units with high weights and vice versa, latent source domains are selected as source spatial units with high weights. TransGTR further selects graph-structured latent source domains by selecting both spatial units and the spatial relations between them, which is achieved by learning transferable graph structures for the selected spatial units in the source city and the target city. Extensive experimental results on real-world datasets show the effectiveness of CrossTReS and TransGTR compared to existing, non-selective baselines. In the future thesis work, we will focus on the sub-problem of selective transfer learning for zero-shot urban STF, which is more challenging as the algorithm should ensure generalization to an unknown distribution of target data. Existing works often perform large-scale pre-training over diverse spatio-temporal data to ensure generalization, which exacerbates the multi-source non-IID issue. Our future thesis work plans to decouple and learn from different latent source domains with mixture-of-experts (MoE) and select adequate experts at test time with a one-shot transferability estimation. Date: Thursday, 5 December 2024 Time: 2:00pm - 4:00pm Venue: Room 3494 Lifts 25/26 Committee Members: Prof. Qiang Yang (Supervisor) Prof. Kai Chen (Supervisor) Dr. Yangqiu Song (Chairperson) Dr. Binhang Yuan