A Multiscale Interactive Recurrent Network for Time-Series Forecasting.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Time-series forecasting is crucial for intelligent applications but faces challenges with complex temporal dependencies.
• Existing multiscale models often fail to explicitly model interrelationships between subseries.

Purpose of the Study:

• To propose a novel multiscale interactive recurrent network (MiRNN) for enhanced time-series forecasting.
• To jointly capture multiscale patterns and model interrelationships between subseries.

Main Methods:

• MiRNN utilizes a deep wavelet decomposition network to create multiscale subseries.
• Key strategies include truncation, initialization, and message passing to model interrelationships.
• A dual-stage attention mechanism captures multiscale temporal dependencies.

Main Results:

• Experiments on four real-world datasets show MiRNN's superior performance.
• The model effectively captures both short-term and long-term temporal dependencies.

Conclusions:

• MiRNN offers a promising approach for time-series forecasting by effectively modeling multiscale patterns and interrelationships.
• The proposed model outperforms existing state-of-the-art methods in forecasting accuracy.