Table of Contents
This repository provides the implementation of the Layer-wise Relevance Propagation (LRP) explanation method for GRU cells (as proposed by the Pytorch framework), as well as for a sequence-to-sequence neural network architecture. We use LRP in order to explain the decisions of an encoder-decoder GRU-based pollution forecasting model.
The steps you need to run to install the required dependencies are the following:
- create environment lrpenv
conda create -n lrpenv python=3.8
- activate environment lrpenv
conda activate lrpenv
- install pip
conda install pip
- install requirements
pip install -r requirements.txt
The folder LRP/ contains the main part of the LRP implementation for a seq-2-seq model with GRU layers.
The folder LRP_toyTask/ contains the scripts used for validation of the LRP implementation through a toy task.
The folder LRP_pollutionForecastModel/ contains the scripts used for applying LRP to a pollution forecasting task.
Sara Mirzavand Borujeni: sara.mirzavand.borujeni@hhi.fraunhofer.de - sarah.mb@outlook.com
Project Link: https://github.com/Sara-mibo/LRP_EncoderDecoder_GRU
- Mirzavand Borujeni, S., Arras, L., Srinivasan, V. et al. Explainable sequence-to-sequence GRU neural network for pollution forecasting. Sci Rep 13, 9940 (2023).
- Petry et al. 2021, Design and Results of an AI-Based Forecasting of Air Pollutants for Smart Cities, ISPRS Ann. Photogramm. Remote Sens. Spatial Inf. Sci., VIII-4/W1-2021, pages 89–96
- Reference Implementation of Layer-wise Relevance Propagation (LRP) for LSTMs, repository by L. Arras
- Arras et al. 2017, Explaining Recurrent Neural Network Predictions in Sentiment Analysis, Proc. of the 8th Work. on Comput. Appr. to Subjectivity, Sentiment and Social Media Analysis, ACL, pages 159-168
- Bach et al. 2015, On Pixel-Wise Explanations for Non-Linear Classifier Decisions by Layer-Wise Relevance Propagation, PLoS ONE 10(7): e0130140