This repository contains the official code implementation for the paper "Towards Principled Task Grouping for Multi-Task Learning" (arXiv:2402.15328).
This paper presents a novel approach to task grouping in Multitask Learning (MTL), advancing beyond existing methods by addressing key theoretical and practical limitations. Unlike prior studies, our approach offers a more theoretically grounded method that does not rely on restrictive assumptions for constructing transfer gains. We also propose a flexible mathematical programming formulation which can accommodate a wide spectrum of resource constraints, thus enhancing its versatility. Experimental results across diverse domains, including computer vision datasets, combinatorial optimization benchmarks and time series tasks, demonstrate the superiority of our method over extensive baselines, validating its effectiveness and general applicability in MTL.
For each dataset, please refer to the original code repository for installation guidance.
Use the following commands to collect required data for task grouping:
cd taskonomy
bash run_collect.shcd celeba/ours_code
python celeba-ours-collect.pycd cop
bash script_collect.shcd ettm1
bash scripts/combinations/longterm_ETTh1_collect.shIf you find this code useful for your research, please cite the original paper:
@misc{wang2024principledtaskgroupingmultitask,
title={Towards Principled Task Grouping for Multi-Task Learning},
author={Chenguang Wang and Xuanhao Pan and Tianshu Yu},
year={2024},
eprint={2402.15328},
archivePrefix={arXiv},
primaryClass={cs.LG},
url={https://arxiv.org/abs/2402.15328},
}This project is licensed under the MIT License. See the LICENSE file for more details.
This repository also includes third-party code from the following projects:
- TAG from Google Research - Licensed under Apache 2.0.
- Taskgrouping by tstandley - Licensed under the MIT License.
- Time-Series-Library by thuml - Licensed under the MIT License.
- MTL-COP by Wastedzz - Licensed under the MIT License.