Disentangled representation learning for spatio-temporal data

Interest in spatio-temporal (ST) prediction tasks has substantially over recent years. In particular, an increasing number of studies has examined the challenging problem of modelling the long-term temporal dependencies of dynamical systems, with applications in domains such as mobility prediction. A key challenge in accurate modelling is learning the relevant representation of ST data. This is particularly important given the need for new approaches to extract better representation from ST data. Disentangled representation learning (DRL) is an unsupervised learning technique that extracts features of a target distribution into narrowly defined variables and keeps them as independent as possible. Compared with traditional methods, such as autoencoder or principal component analysis, disentangled representation can offer insights into the relationships between latent features and semantic attributes as well as providing additional benefits for downstream tasks. In this thesis, we introduce DRL approaches into ST deep learning to extract knowledge from ST data to solve real-world problems. Prior work demonstrates that introducing DRL yields better representation of the input data, thereby boosting downstream task performance for image tasks. However, applying DRL to ST data is still a complex task with some key challenges. These include: (1) Evaluation of the disentanglement with limited or no labels available. Prior research has primarily explored the effectiveness of DRL approaches on image datasets; however, the evaluation metrics often heavily on labels and supervised evaluation approaches. Most ST datasets come without labels and work in an unsupervised manner. (2) Separation of ST features. Most state-of-the-art models use sophisticated mechanisms to extract features from ST domains, but these are often entangled and hard to validate. (3) Usefulness of the learned representations in downstream tasks. Prior work often simply focused on the regression/classification accuracy of the model. However, with the help of DRL methods, the model can separate features (factors of variation) that relate to certain desired properties. Whether DRL can further improve the model’s performance in fairness needs more attention. First, we address the challenge of evaluating the disentanglement of the learned representation from ST data with no label. We apply state-of-the-art DRL methods to ST data and propose an evaluation metric that balances both the disentanglement of the representation and its downstream task utility. We then conduct empirical experiments on multiple datasets to explore whether it is feasible to apply DRL methods to ST data and their effectiveness. Second, we investigate how to separate the spatial and temporal features from ST input and explore insights gained from these features. We propose a novel disentangled learning architecture to further disentangle the representation of ST tasks. This architecture can disentangle the learned latent variables into two sub-groups: time-irrelevant (spatial) features and temporal features. Our experiment results on multiple ST datasets show that the representation learned by our approach has a similar performance to current state-of-the-art methods. Finally, we explore the usefulness of the learned disentangled representations in downstream tasks, not only on the performance of prediction but also on fairness measures. The fairness metric allows us to measure the fairness of predictions made by a neural network related to a certain area on which we focus. We propose a fairness-aware model for ST forecasting, specifically for predicting demand for mobility services. By introducing DRL and adversarial learning, our model can learn to disentangle features comprising sensitive information. We further align multiple real-world ST datasets with a consistent domain, and find that our model can learn generalised representations that can be used for a variety of downstream tasks. In summary, this thesis contributes by applying DRL approaches to ST data and demonstrates how to utilise these approaches to improve model performance for real-world mobility tasks. We hope this study will inspire the research community to develop more robust and effective approaches to solving ST problems in the future.