README.md

Graph-based Time Series Clustering for End-to-End Hierarchical Forecasting (ICML 2024)

This repository contains the code for the reproducibility of the controlled experiments presented in the paper "Graph-based Time Series Clustering for End-to-End Hierarchical Forecasting" (ICML 2024).

Authors: Andrea Cini, Danilo Mandic, Cesare Alippi

---

In a nutshell

HiGP (Hierarchical Graph Predictor) is a graph-based methodology unifying relational and hierarchical inductive biases in the context of deep learning for time series forecasting.

---

Directory structure

The directory is structured as follows:

.
├── config/
├── lib/
├── tsl/
├── conda_env.yaml
├── default_config.yaml
└── experiments/
    └── run_benchmark.py

Datasets

The datasets used in the experiments are provided by the tsl library. The CER-E dataset can be obtained for research purposes following the instructions at this link.

Configuration files

The config directory stores the configuration files used to run the experiments.

Requirements

To solve all dependencies, we recommend using Anaconda and the provided environment configuration by running the command:

conda env create -f conda_env.yml
conda activate higp

A custom version of the tsl library is included in the folder, we suggest to set the PYTHONPATH accordingly.

Experiments

The script used for the experiments in the paper is in the experiments folder.

• run_benchmark.py is used to train and evaluate models on the datasets considered in the study. As an example, to run the HiGP-TTS model on the METR-LA dataset:

   python -m experiments.run_benchmark config=default model=higp_tts dataset=la 
  

Bibtex reference

If you find this code useful please consider citing our paper:

@article{cini2024graph,
title        = {{Graph-based Time Series Clustering for End-to-End Hierarchical Forecasting}},
author       = {Cini, Andrea and Mandic, Danilo and Alippi, Cesare},
journal      = {International Conference on Machine Learning},
year         = 2024
}