论文全名:A systematic comparison study on hyperparameter optimisation of graph neural networks for molecular property prediction
英文是纯手打的!论文原文的summarizing and paraphrasing。可能会出现难以避免的拼写错误和语法错误,若有发现欢迎评论指正!文章偏向于笔记,谨慎食用
目录
2.5. Experimental investigation
2.5.2. Computational cost as a primary consideration
2.5.3. Performance as primary consideration
2.6. Conclusion and future work
1. 省流版
1.1. 心得
(1)这些哥们儿怎么又来了一篇
2. 论文逐段精读
2.1. Abstract
①Molecular datasets are far smaller than other datasets such as image datasets
②They compared TPE and CMA-ES with Random Search (RS)
2.2. Introduction
①简单介绍了一下GNN的应用领域,模型发展,然后说超参数选择会影响结果,盲目选择非常耗时
②TPE and CMA-ES are two SOTA HPO algorithms (the authors think so)
impediment n. 妨碍,障碍;口吃,结巴
2.3. Related work
2.3.1. Random search
①Algorithm of RS
and Bergstra et al. thought RS is the nature baseline of HPO
②RS limits on broad search space and limited computational cost
2.3.2. TPE
①Sequential model-based global optimisation (SMBO) algorithms could solve the cost problem
②Tree-structured Parzen Estimator (TPE) based on SMBO
③详细地介绍了一下TPE和EI,但是这里不多做解释,有需要的小伙伴可以参考这篇论文(虽然我更建议直接去阅读原文啦)
④Algorithm of TPE
2.3.3. CMA-ES
①Algorithm of CAM-ES
which is a derivative-free evolutionary algorithm and can solve black-box optimization problem
2.4. Methodology comparison
2.4.1. Common features
①Randomness
②Derivative-free
③Termination condition
2.4.2. Specific features
①Uniform Distribution vs Gaussian Mixture Model vs Multivariate Normal Distribution
②Model-based and Model free
③Bayesian Optimisation vs Evolutionary Strategy
2.5. Experimental investigation
2.5.1. Experimental settings
①Dataset: ESOL with 1128 samples, FreeSolv with 642 samples and Lipophilicity with 4200 samples
②GNN chosen: GCN
③Hyperparameter selected: batch size , learning rate , size of fully connected layer , size of GCN layer
④Hyperparameter of HPO: default
⑤Training/validation/testing set: 80%/10%/10%
⑥Evaluation metric: root mean square error (RMSE)
⑦Epoch: 30
⑧t-test applied
2.5.2. Computational cost as a primary consideration
①Performance on ESOL:
②Performance on FreeSolv:
③Performance on Lipophicity:
④t-test on 3 datasets:
⑤Time limited test:
⑥t-test on time limited test:
2.5.3. Performance as primary consideration
①Experiments on Repeated HPO Runs on ESOL
②Correspond t-test
③Experiments on A Larger Search Space
②t test:
2.6. Conclusion and future work
①Under these tests, TPE is the best HPO algorithm and RS is the simplest method
②They want to further research the meta hyperparameter
③They hope to cross subjects
3. Reference
Yuan, Y., Wang, W. & Pangm, W. (2021) 'A systematic comparison study on hyperparameter optimisation of graph neural networks for molecular property prediction', Annual Conference on Genetic and Evolutionary Computation. doi: 10.1145/3449639.3459370