• autoregression
• bayesian learning
• generative models
• graph neural networks
• machine learning
• recurrent neural networks
• time series

Traditional generative models for time series often prioritize temporal relationships while ignoring the spatial dependencies intrinsic to many datasets. This thesis proposes an innovative multivariate time series (MTS) generation approach. Our novelty, the Adaptive Spatial Generative Time-series Mixture model (ASGTM), leverages graph-based representations to capture both temporal and spatial dynamics, providing a robust framework for generating synthetic MTS data. ASGTM introduces a self-adaptive spatial component to capture the inter-feature relations in the data where graph-based representations are missing.

Our approach is evaluated against generative methods present in the literature, such as DoppleGANger and Probabilistic Autoregressive, as well as a baseline Mixture Density Network (MDN) across synthetic and real-world datasets.

Results indicate that our proposed models outperform the state of the art in generation quality while also showing competitive performance in prediction and imputation tasks relative to models like Long-Short-Term Memory and Recurrent Neural Networks. This study demonstrates the effectiveness of graph-based generative models in MTS generation, offering a new approach across diverse fields where both temporal and spatial dynamics are essential.