Interpretable dimensionality reduction of multivariate time series data using LSTM based autoencoders

Data collection over time is a common practice in many large organizations- including financial institutions and health care providers- often with the goal of using this data to predict future challenges and opportunities. While this data may contain valuable information, it is often unstructured, coming from different sources and recorded at different times. This lack of structure makes extracting useful information difficult, as most standard statistical and machine learning tools are designed to work with data in a fixed structure. This project will develop a framework for automatically learning a fixed length representation composed of interpretable features from unstructured data collected over time, which requires minimal intervention by human experts. The efficacy of the framework will be evaluated by learning representations for electronic health records, created by the Synthea simulator.

Faculty Supervisor:

Ting Hu;Yuanzhu Chen

Student:

Kyle Nickerson

Partner:

Verafin Inc.

Discipline:

Computer science

Sector:

Information and cultural industries

University:

Memorial University of Newfoundland