Events

The Institute of Artificial Intelligence for Digital Health invites you to attend its monthly seminar series, which will feature Dr. Tengfei Ma, an Assistant Professor in the Department of Biomedical Informatics at Stony Brook University.

Abstract

As one of the fundamental forms of data, time series are prevalent across a wide range of domains and applications, including healthcare, power systems, climate science, and finance. Despite their success in various time series analysis tasks, most deep learning models suffer from an intrinsic drawback: a lack of interpretability. This limitation significantly impedes their real-world applications, particularly in domains such as healthcare, where interpretability is closely linked to reliability and trust. Furthermore, modeling time series data across multiple domains presents additional challenges due to the heterogeneous nature of the data. Different domains can exhibit varying sampling rates, lengths, magnitudes, frequencies, and noise levels. Consequently, most existing machine learning methods for time series modeling tend to focus on single datasets or monotonic domains and struggle to generalize across this diverse landscape.

In this talk, Dr. Ma will introduce some recent works addressing these issues. First, Dr. Ma will discuss how to integrate logic and deep learning in a neuro-symbolic model for time series classification and event stream modeling. The models can automatically discover differentiable logic rules, making predictions readily explainable. Secondly, he will talk about a new approach to discovering latent graphs that connect time series at different locations, enabling unified event detection across diverse data streams. At last, he will discuss recent efforts in developing time series foundation models using shapelets. This approach enables cross-domain pretraining and provides interpretable classifications explained by shapelet patterns. He will also showcase how these techniques can be used in real-world scenarios such as wound healing and mortality prediction.