Multivariate Time Series Information Bottleneck.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Traditional time series (TS) and multiple time series (MTS) prediction models often use distinct deep learning architectures.
• Existing methods for modeling the temporal dimension include decomposition, biologically inspired neural networks, and transformer models.
• The information bottleneck (IB) framework has been underutilized in TS and MTS analyses.

Purpose of the Study:

• To propose a novel deep learning approach for MTS prediction by adapting image processing techniques.
• To investigate the efficacy of compressing the temporal dimension for MTS analysis.
• To introduce the multiple time series-information bottleneck (MTS-IB) model.

Main Methods:

• Encoding time series data into a 2D image-like representation using partial convolution.
• Leveraging advances in image extension for predicting unseen parts of the time series.
• Applying the MTS-IB model to diverse real-world datasets, including electricity production, road traffic, and solar activity.

Main Results:

• The proposed MTS-IB model demonstrates competitive performance compared to traditional TS models.
• The model is grounded in information-theoretical principles.
• The approach is adaptable for higher-dimensional data beyond time and space.

Conclusions:

• The MTS-IB model offers an effective and versatile method for MTS prediction.
• Compressing the temporal dimension is a valuable strategy for MTS analysis.
• The model's applicability spans various domains, including energy, transportation, and astrophysics.