B1682
Title: Transfer learning for time series models
Authors: Abolfazl Safikhani - George Mason University (United States) [presenting]
Abstract: In recent years, the significance of transfer learning has grown substantially due to its ability to swiftly adapt models to novel tasks, environments, and datasets. This adaptation not only enhances the accuracy of machine learning models but also reduces the time needed for training. Nonetheless, the theoretical underpinnings of transfer learning algorithms are somewhat limited, especially in the context of time series models. The primary focus is on high-dimensional vector autoregressive models. A two-step procedure for applying transfer learning is presented, leveraging auxiliary datasets, and subsequently constructing confidence intervals for model parameters in high-dimensional scenarios. Additionally, a novel method is introduced for selecting informative subsets from these auxiliary datasets. The theoretical properties of all the proposed algorithms are established under relatively mild conditions, allowing for dependencies between auxiliary and target datasets. When the informative models exhibit a certain degree of similarity to the target model, the algorithm is proven to achieve the minimax rate. Finally, the empirical performance of the proposed methods is evaluated by analyzing both simulated data and a real-world dataset.
