Still Competitive: Revisiting Recurrent Models for Irregular Time Series Prediction.

Area of Science:

• Machine Learning
• Time Series Analysis
• Recurrent Neural Networks

Background:

• Modeling irregularly sampled multivariate time series presents a significant challenge across various domains, including healthcare and sensor networks.
• Existing complex learning architectures for irregular time series prediction lack clarity regarding their true benefits, prompting investigation into simpler, efficient alternatives.

Purpose of the Study:

• To propose and evaluate GRUwE (Gated Recurrent Unit with Exponential basis functions), an RNN-based algorithm designed for irregularly sampled time series.
• To assess GRUwE's competitiveness against state-of-the-art methods for both regression-based and event-based predictions in continuous time.

Main Methods:

• GRUwE maintains a Markov state representation updated by irregular observations using observation-triggered and time-triggered resets with learnable exponential decays.
• The model supports continuous-time predictions for both next-observation and next-event tasks.

Main Results:

• Empirical evaluations on real-world benchmarks show GRUwE achieving competitive or superior performance compared to current state-of-the-art methods.
• GRUwE demonstrates effectiveness in both next-observation and next-event prediction tasks.

Conclusions:

• GRUwE presents a simple yet powerful RNN-based approach for irregularly sampled time series.
• The model offers practical advantages including ease of implementation, minimal hyper-parameter tuning, and reduced computational overhead for online deployment.