sUrban

Abstract

Recent machine learning research on smart cities has achieved great success in predicting future trends, under the key assumption that the test data follows the same distribution of the training data. The rapid urbanization, however, makes this assumption challenging to hold in practice. Because new data is emerging from new environments (e.g., an emerging city or region), which may follow different distributions from data in existing environments. Different from transfer-learning methods accessing target data during training, we often do not have any prior knowledge about the new environment. Therefore, it is critical to explore a predictive model that can be effectively adapted to unseen new environments. This work aims to address this Out-of-Distribution (OOD) challenge for sustainable cities. We propose to identify two kinds of features that are useful for OOD prediction in each environment: (1) the environment-invariant features to capture the shared commonalities for predictions across different environments; and (2) the environment-aware features to characterize the unique information of each environment. Take bike riding as an example. The bike demands of different cities often follow the same pattern that they significantly increase during the rush hour on workdays. Meanwhile, there are also some local patterns in each city because of different cultures and citizens' travel preferences. We introduce a principled framework -- sUrban -- that consists of an environment-invariant optimization module for learning invariant representation and an environment-aware optimization module for learning environment-aware representation. Evaluation on real-world datasets from various urban application domains corroborates the generalizability of sUrban. This work opens up new avenues to smart city development.