Reproduce “Multivariate LSTM-FCNs for Time Series Classification” on OpenResearch

About this paper

Fazle Karim, Somshubra Majumdar, Houshang Darabi, Samuel Harford

Researchers from the University of Illinois at Chicago developed Multivariate LSTM-FCN (MLSTM-FCN) and Multivariate Attention LSTM-FCN (MALSTM-FCN) models, incorporating Squeeze-and-Excitation blocks, to achieve state-of-the-art performance in multivariate time series classification. These models demonstrated superiority on 28 and 27 out of 35 datasets, respectively, while a dimension shuffle operation reduced processing time by 59% without affecting accuracy.

Key insights

• The Squeeze-and-Excitation (SE) block effectively models and adaptively recalibrates inter-dependencies between variables, proving crucial for enhancing multivariate time series classification performance.
• A novel "dimension shuffle" operation can reduce total processing time by 59% across datasets without statistically altering classification accuracy, demonstrating significant practical efficiency gains.
• Hybrid CNN-LSTM architectures, augmented with channel-wise attention mechanisms like the SE block, offer a powerful and generalizable approach for diverse multivariate time series tasks.

Summary by alphaXiv.