Interpreting Multi-Horizon Time Series Deep Learning Models

Interpreting the model's behavior is important in understanding decision-making in practice. However, explaining complex time series forecasting models faces challenges due to temporal dependencies between subsequent time steps and the varying importance of input features over time. Many time series forecasting models use input context with a look-back window for better prediction performance. However, the existing studies (1) do not consider the temporal dependencies among the feature vectors in the input window and (2) separately consider the time dimension that the feature dimension when calculating the importance scores. In this work, we propose a novel Windowed Temporal Saliency Rescaling method to address these issues.

Core Libraries

The following libraries are used as a core in this framework.

Captum

(“comprehension” in Latin) is an open source library for model interpretability built on PyTorch.

Time Interpret (tint)

Expands the Captum library with a specific focus on time-series. It includes various interpretability methods specifically designed to handle time series data.

Time-Series-Library (TSlib)

TSlib is an open-source library for deep learning researchers, especially deep time series analysis.

Interpretation Methods

The following local intepretation methods are supported at present:

1. *Feature Ablation* [[2017]](https://arxiv.org/abs/1705.08498) 2. *Dyna Mask* [[ICML 2021]](https://arxiv.org/abs/2106.05303) 3. *Extremal Mask* [[ICML 2023]](https://proceedings.mlr.press/v202/enguehard23a/enguehard23a.pdf) 4. *Feature Permutation* [[Molnar 2020]](https://christophm.github.io/interpretable-ml-book/) 5. *Augmented Feature Occlusion* [[NeurIPS 2020]](https://proceedings.neurips.cc/paper/2020/file/08fa43588c2571ade19bc0fa5936e028-Paper.pdf) 6. *Gradient Shap* [[NeurIPS 2017]](https://proceedings.neurips.cc/paper/2017/file/8a20a8621978632d76c43dfd28b67767-Paper.pdf) 7. *Integreated Gradients* [[ICML 2017]](https://proceedings.mlr.press/v70/sundararajan17a/sundararajan17a.pdf) 8. *WinIT* [[ICLR 2023 poster]](https://openreview.net/forum?id=C0q9oBc3n4) 9. *TSR* [[NeurIPS]](https://proceedings.neurips.cc/paper_files/paper/2020/file/47a3893cc405396a5c30d91320572d6d-Paper.pdf) 10. *WinTSR* - proposed new method

Time Series Models

This repository currently supports the following models:

• TimeMixer - TimeMixer: Decomposable Multiscale Mixing for Time Series Forecasting [ICLR 2024] [Code].
• TSMixer - TSMixer: An All-MLP Architecture for Time Series Forecasting [arXiv 2023] [Code]
• TimesNet - TimesNet: Temporal 2D-Variation Modeling for General Time Series Analysis [ICLR 2023] [Code]
• DLinear - Are Transformers Effective for Time Series Forecasting? [AAAI 2023] [Code]
• PatchTST - A Time Series is Worth 64 Words: Long-term Forecasting with Transformers. [ICLR 2023] [Code]
• MICN - MICN: Multi-scale Local and Global Context Modeling for Long-term Series Forecasting [ICLR 2023][Code]
• Crossformer - Crossformer: Transformer Utilizing Cross-Dimension Dependency for Multivariate Time Series Forecasting [ICLR 2023][Code]
• SegRNN - SegRNN: Segment Recurrent Neural Network for Long-Term Time Series Forecasting [arXiv 2023] [Code].
• Koopa - Koopa: Learning Non-stationary Time Series Dynamics with Koopman Predictors [NeurIPS 2023] [Code].
• FreTS - Frequency-domain MLPs are More Effective Learners in Time Series Forecasting [NeurIPS 2023] [Code].
• TiDE - Long-term Forecasting with TiDE: Time-series Dense Encoder [arXiv 2023] [Code].
• LightTS - Less Is More: Fast Multivariate Time Series Forecasting with Light Sampling-oriented MLP Structures [arXiv 2022] [Code]
• ETSformer - ETSformer: Exponential Smoothing Transformers for Time-series Forecasting [arXiv 2022] [Code]
• Non-stationary Transformer - Non-stationary Transformers: Exploring the Stationarity in Time Series Forecasting [NeurIPS 2022] [Code]
• FEDformer - FEDformer: Frequency Enhanced Decomposed Transformer for Long-term Series Forecasting [ICML 2022] [Code]
• Pyraformer - Pyraformer: Low-complexity Pyramidal Attention for Long-range Time Series Modeling and Forecasting [ICLR 2022] [Code]
• FiLM - FiLM: Frequency improved Legendre Memory Model for Long-term Time Series Forecasting [NeurIPS 2022][Code]
• Autoformer - Autoformer: Decomposition Transformers with Auto-Correlation for Long-Term Series Forecasting [NeurIPS 2021] [Code]
• Informer - Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting [AAAI 2021] [Code]
• Reformer - Reformer: The Efficient Transformer [ICLR 2020] [Code]
• Transformer - Attention is All You Need [NeurIPS 2017] [Code]

Train & Test

Use the run.py script to train and test the time series models. Check the scripts and slurm folder to see sample scripts. Make sure you have the datasets downloaded in the dataset folder following the Datasets section. Following is a sample code to train the electricity dataset using the DLinear model. To test an already trained model, just remove the --train parameter.

python run.py \
  --task_name long_term_forecast \
  --train \
  --root_path ./dataset/electricity/ \
  --data_path electricity.csv \
  --model DLinear \
  --features S \
  --seq_len 96 \
  --label_len 12 \
  --pred_len 24 \
  --n_features 1

Interpret

Use the interpret.py script to interpret a trained model. Check the scripts and slurm folder to see more sample scripts. Following is a sample code to interpret the iTransformer model trained on the electricity dataset using using some of the interpretation methods. This evaluates the 1st iteration among the default 3 in the result folder.

python interpret.py \
  --task_name long_term_forecast \
  --explainers feature_ablation augmented_occlusion feature_permutation integrated_gradients gradient_shap wtsr\
  --root_path ./dataset/electricity/ \
  --data_path electricity.csv \
  --model iTransformer \
  --features S \
  --seq_len 96 \
  --label_len 12 \
  --pred_len 24 \
  --n_features 1 \
  --itr_no 1

Datasets

The datasets are available at this Google Drive in the long-term-forecast folder. Download and keep them in the dataset folder here. Only mimic-iii dataset is private and hence must be approved to get access from PhysioNet.

Electricity

The electricity dataset ¹ was collected in 15-minute intervals from 2011 to 2014. We select the records from 2012 to 2014 since many zero values exist in 2011. The processed dataset contains the hourly electricity consumption of 321 clients. We use ’MT 321’ as the target, and the train/val/test is 12/2/2 months. We aggregated it to 1h intervals following prior works.

Traffic

This dataset ² records the road occupancy rates from different sensors on San Francisco freeways.

Mimic-III

MIMIC-III is a multivariate clinical time series dataset with a range of vital and lab measurements taken over time for around 40,000 patients at the Beth Israel Deaconess Medical Center in Boston, MA (Johnson et al. ³, 2016). It is widely used in healthcare and medical AI-related research. There are multiple tasks associated, including mortality, length-of-stay prediction, and phenotyping.

We follow the pre-processing procedure described in Tonekaboni et al. (2020) [^4] and use 8 vitals and 20 lab measurements hourly over a 48-hour period to predict patient mortality. For more visit the [source description](https://physionet.org/content/mimiciii/1.4/).

This is a private dataset. Refer to the official MIMIC-III documentation. ReadMe and datagen of MIMIC is from Dynamask Repo. This repository followed the database setup instructions from the offficial site here.

• Run this command to acquire the data and store it:

  python -m data.mimic_iii.icu_mortality --sqluser YOUR_USER --sqlpass YOUR_PASSWORD

  If everything happens properly, two files named adult_icu_vital.gz and adult_icu_lab.gz are stored in dataset/mimic_iii.

• Run this command to preprocess the data:

  python -m data.mimic_iii.data_preprocess

  If everything happens properly, a file mimic_iii.pkl is stored in dataset/mimic_iii.

Reproduce

The module was developed using python 3.10.

Option 1. Use Container

Dockerfile contains the docker buidling definitions. You can build the container using

docker build -t timeseries

This creates a docker container with name tag timeseries. The run our scripts inside the container. To create a Singularity container use the following definition file.

sudo singularity build timeseries.sif singularity.def

This will create a singularity container with name timeseries.sif. Note that, this requires sudo privilege.

Option 2. Use Virtual Environment

First create a virtual environment with the required libraries. For example, to create an venv named ml, you can either use the Anaconda library or your locally installed python. An example code using Anaconda,

conda create -n ml python=3.10
conda activate ml

This will activate the venv ml. Install the required libraries,

python3 -m pip install -r requirements.txt

If you want to run code on your GPU, you need to have CUDA installed. Check if you already have CUDA installed.

import torch

device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
print(f'Using {device} backend')

If this fails to detect your GPU, install CUDA using,

pip install torch==2.2 --index-url https://download.pytorch.org/whl/cu118

References

Footnotes

1. https://archive.ics.uci.edu/ml/datasets/ElectricityLoadDiagrams20112014. ↩

2. https://pems.dot.ca.gov/. ↩

3. Alistair EW Johnson, Tom J Pollard, Lu Shen, Li-wei H Lehman, Mengling Feng, Mohammad Ghassemi, Benjamin Moody, Peter Szolovits, Leo Anthony Celi, and Roger G Mark. MIMIC-III, a freely accessible critical care database. Scientific Data, 3, 2016. ↩