High-Order Temporal Correlation Model Learning for Time-Series Prediction.

Area of Science:

• Data Science
• Machine Learning
• Signal Processing

Background:

• Time-series prediction is challenging with multiway array data due to noise and redundancy.
• Existing methods may not fully capture the intricate temporal structures in high-order tensor representations.

Purpose of the Study:

• To develop novel multilinear models for high-order time-series prediction using tensor decomposition.
• To address noise and redundancy in tensor representations of time series data.

Main Methods:

• Introduced the multilinear orthogonal autoregressive (MOAR) model, emphasizing orthogonal constraints for information preservation.
• Developed the multilinear constrained autoregressive (MCAR) model, enhancing MOAR with an inverse decomposition error term.
• Projected original tensors into subspaces and generalized autoregressive models to tensor form for temporal smoothness.

Main Results:

• Both MOAR and MCAR models demonstrated convergence within few iterations during training.
• The proposed methods achieved promising prediction performance compared to state-of-the-art techniques.
• Experiments on four public datasets validated the effectiveness of the developed models.

Conclusions:

• The novel MOAR and MCAR models offer effective solutions for high-order time-series prediction.
• Tensor decomposition provides a robust framework for handling complex time-series data.
• The proposed methods preserve temporal smoothness and intrinsic structures for accurate predictions.