2024 2023 2022 2020 2019 2017 2016 2015 2014 2013 2012 2011 2010 

Predictive and Generative Modeling of Extremes in Time Series using Deep Neural Networks

Author: Asadullah Galib
Event Date: 2024-04-03
Location: EB 2250
The accurate modeling of extreme values in time series data is a critical yet challenging task that has garnered significant interest in recent years. The impact of extreme events on human and natural systems underscores the need for effective and reliable modeling methods. Addressing this challenge, I present several novel deep learning frameworks aimed at effectively capturing extreme events in time series data. This talk focuses on two key works from the proposed frameworks. Firstly, DeepExtrema introduces a predictive modeling approach that integrates extreme value theory (EVT) with deep learning to enhance forecast accuracy and reliability. By incorporating EVT principles into deep learning, this approach captures the tail distribution of time series data and provides robust uncertainty estimation. The primary challenge lies in effectively integrating EVT into the deep learning formulation. Secondly, Diffmaxima introduces a conditional diffusion model tailored to address the preservation of extreme value distributions within generative modeling. It combats the premature fade out of local patterns through a high-frequency inflation strategy while extending traditional diffusion-based models to encompass extreme events by generating samples conditioned on extreme value. Evaluation of these frameworks using real-world and synthetic datasets demonstrates their superior performance compared to existing state-of-the-art methods.