Region Profile Enhanced Urban Spatio-Temporal Prediction via Adaptive Meta-Learning

Abstract

Urban spatio-temporal (ST) prediction plays a crucial role in smart city construction. Due to the high cost of ST data collection, improving ST prediction in a lack of data is significant. For this purpose, existing meta-learning methods have been demonstrated powerful by learning an initial network from training tasks and adjusting to target tasks with limited data. However, such shared knowledge from a set of tasks may contain irrelevant noise due to the gap of region-varying ST dynamics, resulting in the negative transfer issue. As a revelation of regional functional patterns, region profiles give rise to the diversity of ST dynamics. Thus, we design a novel adaptive meta-optimized model MetaRSTP, which conducts the initial prediction model in a finer-granularity of region level with region profiles as semantic evidence. To enhance the expressiveness of profiles, we firstly build a semantic alignment space to explore the inter-view co-semantics. Fusing it with view-specific uniqueness, the multi-view region profiles can be better applied in urban tasks. Then, a regional bias generator derives non-shared parameters in terms of profiles, which alleviates the divergence among regions. We set a new meta-learning strategy as initialize the network with fixed generalizable parameters and region-adaptive bias, thus enhancing the personalized prediction performance even in few-shot scenarios. Extensive experiments on real-world datasets illustrate the effectiveness of our MetaRSTP and our learned region profiles.