Part Overview
Forecasting asks a model to predict the next value; representation learning asks it to understand the series well enough that prediction, classification, and retrieval all become easy. Part IV is about that shift in goal. Its three chapters cover the unsupervised and generative side of temporal deep learning: learning embeddings without labels, generating realistic synthetic series, and modeling the irregular event streams and evolving graphs that do not fit a fixed-rate grid. These are the techniques that matter when labels are scarce, when data cannot be shared, or when the signal arrives as discrete events rather than samples.
Chapter 16 develops temporal representation learning proper: self-supervised pretext tasks, contrastive methods such as TS2Vec and TF-C, and masked modeling, ending with embeddings that disentangle the factors behind a series. Chapter 17 turns to generation, walking from sequence autoencoders and VAEs through TimeGAN to the diffusion models (TimeGrad, CSDI) that now lead synthetic temporal data, and it treats evaluation as a first-class problem because a generator is only as good as the tests it passes. Chapter 18 leaves the regular grid entirely for event and sequence modeling: marked event streams, temporal point processes such as Hawkes and neural TPPs, dynamic and temporal graphs, and time-to-event prediction.
The temporal thread continues through this part: the HMM and neural latent-variable model introduced earlier return as the sequential VAE of Chapter 17, and the irregular healthcare running dataset reappears in the time-to-event modeling of Chapter 18, connecting back to the Neural CDE of Part III. The representations and generators built here also feed forward: self-supervised embeddings underpin the foundation-model thinking of Part III, and synthetic generation supports the privacy and data-engine concerns of Part VIII. Part IV is compact but pivotal, the part where temporal models stop merely predicting and start understanding.
Self-supervised pretext tasks, contrastive learning, and masked modeling that yield reusable temporal embeddings without labels.
Sequence autoencoders, VAEs, GANs, and diffusion models for synthesizing realistic series, with evaluation taken seriously.
Marked event streams, temporal point processes, dynamic graphs, and time-to-event modeling beyond the fixed-rate grid.
Where This Part Leads
With representations learned and series generated, the book confronts what every deployed forecaster must answer for: how sure is it, and how does it cope when the world changes. Part V: Uncertainty, Online, and Adaptive Learning builds on these models by wrapping them in honest uncertainty quantification and the online, continual, and adaptive machinery that keeps them accurate under drift.