TSCMamba: Mamba Meets Multi-View Learning for Time Series Classification.

Area of Science:

• Machine Learning
• Data Science
• Signal Processing

Background:

• Multivariate time series classification (TSC) is crucial for healthcare and finance.
• Existing TSC methods largely overlook properties like shift equivariance and inversion invariance.
• There is a need for robust TSC models that capture complex temporal patterns.

Purpose of the Study:

• To propose a novel multi-view approach for TSC that incorporates shift equivariance and inversion invariance.
• To enhance the accuracy and robustness of time series classification models.
• To leverage diverse feature representations and advanced sequence modeling for improved TSC.

Main Methods:

• A multi-view approach integrating spectral, temporal, local, and global features.
• Continuous wavelet transform for time-frequency analysis and shift-consistent features.
• Fusion of features with temporal convolutional or multilayer perceptron networks.
• Mamba state space model with a novel 'tango scanning' scheme for sequence modeling and inversion invariance.
• Experiments on benchmark datasets (10+20) for performance evaluation.

Main Results:

• Achieved average accuracy improvements of 4.01-6.45% and 7.93% over leading TSC models.
• Demonstrated effectiveness in capturing patterns with shift equivariance and inversion invariance.
• The proposed method shows enhanced generalization and robustness compared to TimesNet and TSLANet.

Conclusions:

• The novel multi-view approach significantly improves multivariate time series classification.
• Incorporating shift equivariance and inversion invariance enhances model performance.
• The Mamba model with tango scanning offers an efficient and scalable solution for TSC.