A1196
Title: Transformer-based models for time series forecasting
Authors: Thu Nguyen - University of Maryland Baltimore County (United States) [presenting]
Abstract: Time series forecasting has long been a key and well-studied area of academic research and has many important applications in topics such as commercial decision-making in retail, finance, product development and planning, biological sciences and medicine, to name a few. In contrast to traditional methods for time series forecasting, which focus on parametric models informed by domain expertise and rely heavily on well-designed features, modern machine learning methods, especially deep learning, which gained popularity in recent times, try to learn temporal dynamics in a purely data-driven manner. With the increasing data availability and computing power in recent times, as well as their success in practical applications and competitions, deep learning-based forecasting models have become a popular choice for many time series forecasting tasks. We present a new class of models inspired by Transformer based architecture for the time series forecasting tasks. The proposed models' architecture is based on the attention mechanism, which has received increased interest in applying for time series-related applications. In addition to employing new attention mechanisms, we also utilize some ideas from classical time series methods to learn complex patterns and dynamics from time series data. The proposed models are flexible enough to apply to both univariate and multivariate time series data.
