Fahim Tasneema Azad - PhD student Data Science
Date: - Location: Eurecom
Abstract: Numerous IoT-enabled sensors generate long sequence data in complex systems. Technological advancement enables us to retrieve this complex multivariate time series data from these systems. Therefore, developing a situational understanding and the ability to make informed decisions based on this complex data is critical for operating effectively, efficiently, safely, and securely in communities and human-centered environments. Time series forecasting and causal inference are essential in decision-making in such a complex system. We aim to develop techniques to forecast multivariate time series and perform a causal analysis that will enable users to assess the situation and make informed decisions in a complex system in the real world. To achieve these goals, we propose to leverage the attention mechanism to forecast and draw causal inferences. We develop novel attention mechanisms that enable machine learning models to predict based on long sequence data from a relatively small set of sampled data. The initial time complexity calculations show that such techniques would reduce the computation time for a long sequence without affecting accuracy. In this talk, we present machine learning models to forecast time series using attention mechanisms in the context of the pandemic and perform causal inference based on the time series data. Short bio: Fahim Tasneema Azad is a Ph.D. student in Computer Science at the School of Computing and Augmented Intelligence at Arizona State University, working with Dr. K. Selçuk Candan. She has earned her Master's in Computer Science from the same institution with distinction. Her primary research interests lie in Time Series Analysis, Machine Learning, Causal Inference, and Data Science. Her research focuses on causally-informed multivariate time series forecasting in complex coupled systems. Currently, she is investigating the effectiveness of incorporating approximate attention mechanisms in transformers to forecast time series data.
