PhD Dissertation Proposal by Gerald WOO | Deep Representation Learning for Time Series Forecasting

Time series forecasting is an important task in many business and scientific applications -- from demand forecasting for inventory management to financial time series prediction for portfolio management. The advent of deep learning methodologies in the past decade has paved a path forward for the next generation of time series forecasting methods. Deep representation learning for time series forecasting presents an opportunity to design more flexible models and learning paradigms for time series forecasting, allowing us to learn time series representations from data -- allowing for unparalleled scaling in terms of dataset size and model size, something inaccessible in the classical setting. At the same time, the deep learning paradigm allows us to retain a fine-grained control over specifying inductive biases into our models. In this dissertation proposal, we present two threads of work. Firstly, inspired by classical time series analysis methods, we explore how to apply the inductive biases of seasonality and trend into both neural network architectures, as well as the learning paradigm for these models. Secondly, we further explore a new class of time series models, known as deep time-index models. These models are the deep learning analogue of time series regression models on time-index features. We show that learning these models naively leads to overfitting on the training set and present an effective approach to learn these models. Finally, we look towards the future, presenting our plans to further explore probabilistic deep time-index models, and scaling deep time series forecasting models.

Gerald Woo is a Ph.D. candidate at the School of Computing and Information Systems, Singapore Management University. He is advised by Assoc. Prof. Akshat Kumar. He is part of the Industrial Postgraduate Programme with Salesforce Research Asia, advised by Chenghao Liu (Senior Applied Scientist). His research interest lies in the domain of time series and is working on neural network models and representation learning frameworks for time series forecasting.