Multi-task Graph Neural Architecture Search with Task-aware Collaboration and Curriculum

Authors: Yijian Qin, Xin Wang, Ziwei Zhang, Hong Chen, Wenwu Zhu

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments on both synthetic and real-world datasets validate the superiority of our proposed MTGC3 model over existing baselines via customizing the optimal architecture for each task and sharing useful information among them.
Researcher Affiliation Academia Yijian Qin1,2, Xin Wang1,2 , Ziwei Zhang1, Hong Chen1, Wenwu Zhu1,2 1Department of Computer Science and Technology, Tsinghua University, 2BNRist qinyj19@mails.tsinghua.edu.cn, {xin_wang,zwzhang}@tsinghua.edu.cn h-chen20@mails.tsinghua.edu.cn, wwzhu@tsinghua.edu.cn
Pseudocode Yes Algorithm 1: MTGC3
Open Source Code Yes 3https://github.com/THUMNLab/Auto GL-light
Open Datasets Yes Real-world Datasets. We choose three widely-used multi-task graph classification datasets included OGB [13]: OGBG-Tox21 [14], OGBG-Tox Cast [35], and OGBG-Sider [18].
Dataset Splits Yes We choose three widely-used multi-task graph classification datasets included OGB [13]: OGBG-Tox21 [14], OGBG-Tox Cast [35], and OGBG-Sider [18].
Hardware Specification No The paper does not provide specific details on the hardware used for experiments, such as GPU or CPU models.
Software Dependencies No The paper does not list specific software dependencies with their version numbers.
Experiment Setup Yes Experimental Details. We set the number of layers as 3 for synthetic datasets, and 5 for real-world datasets. For all datasets except Tox Cast, we use the task-separate head. For Tox Cast, we use the cross-mixed head with 16 chunks. We train our models with a batch size of 32 for synthetic datasets and 128 for real-world datasets. We use Adam optimizer with a learning rate of 0.001.