A New Auto-Regressive Multi-Variable Modified Auto-Encoder for Multivariate Time-Series Prediction: A Case Study with...

Area of Science:

• Epidemiology
• Data Science
• Artificial Intelligence

Background:

• The COVID-19 pandemic highlighted challenges in accurate forecasting due to data limitations.
• Existing epidemiological and machine learning models showed effectiveness but with precision constraints for multi-variable pandemic data.

Purpose of the Study:

• To propose and evaluate a novel stacked auto-encoder approach for enhanced time-series prediction.
• To address limitations in multi-variable forecasting for pandemic scenarios.

Main Methods:

• Developed a novel time-series prediction approach using stacked auto-encoder structures with variations for training and weight adjustment.
• Conducted comparative experiments using COVID-19 case data, environmental factors (temperature, humidity, AQI), and global case percentages.

Main Results:

• Achieved an 80.7% decrease in Root Mean Square Error (RMSE) for entire data and a 10.3% decrease for test data across 50 trial-trained models.
• A specific model variation (model type#3) ranked 4th overall, outperforming established models like NBEATS, Prophet, and Glounts.

Conclusions:

• The proposed stacked auto-encoder model demonstrates significant forecasting capacity and versatility.
• This approach shows promise for various time-series tasks, particularly in complex scenarios like pandemic prediction.