Reinforcement Learning-Guided Channel Selection Across Time for Multivariate Time Series Classification
This repository contains the official implementation of the RELEVANT framework described in Reinforcement Learning-Guided Channel Selection Across Time for Multivariate Time Series Classification, accepted at SSCI 2023.
The code is written in Python 3.8.13 and uses the following main dependencies:
- torch==1.13.0
- tsai==0.3.4
- sktime==0.13.4
- numpy==1.21.5
- scikit-learn==1.1.3
- pandas==1.5.1
The initial structure of the framework and files is based on the code for Stop&Hop, an early classification method for irregular time series, by Hartvigsen et al.
The models are evaluated on a subset of the UEA multivariate dataset collection. Moreover, we utilized datasets based on the MAFAULDA Machinery Fault Database and the Case Western Reserve University Bearing Data.
The synthetic dataset used is included in the data directory.