Real-world time series data is often complex and challenging for traditional analysis methods. Deep learning offers promising capabilities for handling this complexity, but effective methods for time series data are still being developed. In particular, the commonly-used sequence-to-sequence framework does not utilize the temporal hierarchical structure and does not perform well in the decoding of long sequences. This thesis explores advanced deep-learning techniques by incorporating temporal hierarchical structure for more effective time series analysis.

First, we propose a recurrent autoencoder with multiresolution ensemble decoding. This uses a coarse-to-fine fusion mechanism for multiresolution temporal fusion of multiple decoder experts. On the other hand, one-class representation...[ Read more ]

Real-world time series data is often complex and challenging for traditional analysis methods. Deep learning offers promising capabilities for handling this complexity, but effective methods for time series data are still being developed. In particular, the commonly-used sequence-to-sequence framework does not utilize the temporal hierarchical structure and does not perform well in the decoding of long sequences. This thesis explores advanced deep-learning techniques by incorporating temporal hierarchical structure for more effective time series analysis.

First, we propose a recurrent autoencoder with multiresolution ensemble decoding. This uses a coarse-to-fine fusion mechanism for multiresolution temporal fusion of multiple decoder experts. On the other hand, one-class representation learning avoids sequential decoding by assuming one hypersphere in the high-dimensional hidden space. We extend existing one-class methods by introducing a temporal hierarchical one-class representation with multiple hyperspheres in multiple temporal scales. This allows capturing abundant normal patterns in a unified one-class learning framework. Additionally, we also explore an adaptive multiple-round masking-based contrastive learning framework for improved one-class representation learning.

Temporal hierarchical structure can also be used in time series diffusion models. We first develop two conditional generation mechanisms to enhance these diffusion models. We then propose to utilize multi-resolution analysis to sequentially extract fine-to-coarse trends from the time series for forward diffusion, with the denoising process proceeding in an easy-to-hard non-autoregressive manner.

Experimental results show that integrating temporal hierarchical information into deep learning models is more effective than state-of-the-art methods across various time series tasks, including normality modeling and generative forecasting.