Abstract:Multivariate time series classification (MTSC) is a crucial machine learning problem prevalent across various real-life domains. Traditional deep learning approaches with high accuracy in MTSC are often criticized for their “black box” nature, offering no insight into their operational mechanisms or decision-making processes. Explainable artificial intelligence (XAI) becomes a key idea for dealing with such limitations in decision-sensitive areas. While some researchers have explored the interpretability of MTSC, the majority have focused on elucidating the relationship between variables over time, neglecting the intricate connections among different variables. To address the interpretability issue in MTSC, we introduce an explainable dual-mode convolutional neural network (XDM-CNN) designed specifically for MTSC. The proposed XDM-CNN framework comprises two modules: a classification module and an explanation module. The classification module can ensure exceptional classification performance by combining both one-dimensional and two-dimensional convolutional neural networks, while the explanation module can mine the underlying logical relationships among variables during the classification process by utilizing a synergy of visualization and quantification techniques. We validate our approach using 26 datasets sourced from the University of East Anglia (UEA) archive. The experimental results demonstrate that XDM-CNN not only exhibits excellent performance in classification accuracy, but also has strong explainability. By combining visual and numerical explanation, the hidden logical relationships among multivariate time series are explored and interpreted, providing human users with a decidable basis for classification decisions.

Multi-Scale Convolutional Neural Networks for Time Series Classification

Multi-scale Attention Convolutional Neural Network for time series classification

Dynamic Multi-Scale Convolutional Neural Network for Time Series Classification.

Exploiting Multi-Channels Deep Convolutional Neural Networks for Multivariate Time Series Classification

XCM: an Explainable Convolutional Neural Network for Multivariate Time Series Classification.

LSTM-MFCN: A time series classifier based on multi-scale spatial–temporal features

Combining contextual neural networks for time series classification

A Deep Learning Framework for Time Series Classification Using Relative Position Matrix and Convolutional Neural Network.

A new multi-process collaborative architecture for time series classification

Attention-Based Deep Gated Fully Convolutional End-to-End Architectures for Time Series Classification

An Explainable Dual-Mode Convolutional Neural Network for Multivariate Time Series Classification

Self-Attention Causal Dilated Convolutional Neural Network for Multivariate Time Series Classification and Its Application

End-to-end Multivariate Time Series Classification Via Hybrid Deep Learning Architectures

Encoding Time Series as Multi-Scale Signed Recurrence Plots for Classification Using Fully Convolutional Networks

Adaptive Multi-Scale Wavelet Neural Network for Time Series Classification

Time Series Classification Based on Multi-scale Dynamic Convolutional Features and Distance Features

MTMI-DCNN: A PSR-Based Method for Time Series Sensor Data Classification

A deep learning framework for time series classification using normal cloud representation and convolutional neural network optimization

FormerTime: Hierarchical Multi-Scale Representations for Multivariate Time Series Classification

Research on time series classification based on convolutional neural network with attention mechanism

Multitask Learning for Time Series Data with 2D Convolution