Please join us for CIRES HDR, Hrishi Patel’s, PhD confirmation milestone seminar.

Self-Supervised Learning for Irregular Time Series

Speaker: Hrishikesh Patel

Abstract : Irregular time series data, common in domains such as finance, Internet-of-Things (IoT), and healthcare, present unique challenges due to their non-uniform recording intervals and asynchronous nature. Traditional models, designed for uniform time intervals, are ill-equipped to handle these irregular datasets.  This issue is exacerbated in healthcare, where the scarcity and high cost of labelled data impose additional constraints. Our primary research question is how we can effectively utilize limited labelled data for irregular time series modelling? To address this, we have turned to self-supervised learning (SSL), which reduces reliance on labelled data by employing pretext tasks tailored for irregular time series. We introduce a novel framework, Event-based Masking for Irregular Time series (EMIT), which enhances data representation by masking and reconstructing embeddings of critical points in latent space. By identifying these critical points using the rate of change, our method ensures that the model focuses on the most informative aspects of the data. Our approach has demonstrated promising results on benchmark healthcare datasets such as MIMIC-III and the PhysioNet Challenge 2012. These results highlight the enhanced ability of our model to capture complex patterns and make accurate predictions, which are crucial for effective healthcare applications.

Bio: Hrishikesh Patel is a PhD candidate at the ARC Centre for Information Resilience (CIRES) under School of EECS, UQ. He received his B.Tech in Petroleum Engineering from the Pandit Deendayal Energy University, India, and master’s in data science from the UQ. His primary research interests are Time Series Modelling, Medical AI, and Self-Supervised Learning.