Recurrent Neural Networks for Multivariate Time Series with Missing Values.

Area of Science:

• Data Science
• Machine Learning
• Artificial Intelligence

Background:

• Multivariate time series data frequently contain missing values across various domains like healthcare and biology.
• Missing data patterns can be correlated with target labels, a phenomenon known as informative missingness.
• Existing methods have limitedly explored exploiting these missing patterns for imputation and prediction enhancement.

Purpose of the Study:

• To develop novel deep learning models that leverage missing data patterns for improved time series analysis.
• To introduce GRU-D, a model based on Gated Recurrent Unit (GRU), designed to incorporate missingness information.
• To enhance the performance of time series prediction tasks by effectively utilizing informative missingness.

Main Methods:

• Developed GRU-D, a deep learning model integrating Gated Recurrent Unit (GRU) architecture.
• Incorporated two representations of missing patterns: masking and time interval, into the model.
• Designed the model to capture long-term temporal dependencies while utilizing missingness characteristics.

Main Results:

• GRU-D achieved state-of-the-art performance on time series classification tasks.
• Experiments were conducted on real-world clinical datasets (MIMIC-III, PhysioNet) and synthetic datasets.
• The model demonstrated the effectiveness of utilizing missing patterns for better prediction.

Conclusions:

• GRU-D offers a novel approach to handling missing data in multivariate time series.
• The findings highlight the importance of understanding and utilizing missingness patterns in time series analysis.
• The model provides a foundation for future research in informative missingness and time series prediction.