Multivariate time-series classification with hierarchical variational graph pooling.

Area of Science:

• Machine Learning
• Data Science
• Artificial Intelligence

Background:

• Advancements in sensing technology have increased the need for effective multivariate time-series classification (MTSC).
• Current deep learning methods for MTSC often overlook complex inter-variable dependencies and lack hierarchical feature aggregation.
• Graph Neural Networks (GNNs) offer potential for modeling variable interactions but existing spatial-temporal methods are typically flat.

Purpose of the Study:

• To propose a novel graph-pooling framework, MTPool, for enhanced multivariate time-series classification.
• To address the limitations of existing methods in capturing hierarchical spatial-temporal dependencies.
• To develop a method that can learn expressive global representations of multivariate time series.

Main Methods:

• MTS data slices are converted into graphs using a graph structure learning module.
• Spatial-temporal node features are extracted using a temporal convolutional module.
• An encoder-decoder based variational graph pooling module is employed for adaptive graph coarsening and representation learning, combined with GNNs.

Main Results:

• The proposed MTPool framework successfully learns expressive global representations of multivariate time series.
• Experiments on ten benchmark datasets demonstrate MTPool's superior performance compared to state-of-the-art MTSC strategies.
• The method effectively captures complex pairwise dependencies among multivariate variables through graph representation.

Conclusions:

• MTPool provides a powerful and effective approach for multivariate time-series classification.
• The hierarchical graph pooling mechanism is crucial for improving classification accuracy.
• This framework offers a promising direction for future research in spatial-temporal data analysis.