[论文精读]Hyperparameter Optimization of Graph Neural Networks for mRNA Degradation Prediction

论文原文：Hyperparameter Optimization of Graph Neural Networks for mRNA Degradation Prediction | IEEE Conference Publication | IEEE Xplore

代码：https://github.com/ViktorijaVodilovska/mRNA_pre

数据：OpenVaccine: COVID-19 mRNA Vaccine Degradation Prediction | Kaggle

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

1. 省流版

1.1. 心得

2. 论文逐段精读

2.1. Abstract

2.2. Introduction

2.3. Related work

2.4. Materials and methods

2.4.1. Dataset

2.4.2. Graph neural networks models

2.4.3. Optimization algorithms

2.5. Experiments, Results and Discussion

2.5.1. Experimental details

2.5.2. Results and discussion

2.6. Conclusion

3. Reference

1. 省流版

1.1. 心得

（1）可以随随便便简单读读，他们就是做了个超参数优化方法测试

2. 论文逐段精读

2.1. Abstract

        ①Contained hyper paramer optimiztion (HPO) algorithms: bio-inspired Genetic Algorithm, Particle Swarm Optimization, and Artificial Bee Colony

pharmaceutical  adj.制药的；配药的；卖药的  n.药物

2.2. Introduction

        ①Grid search for HPO is expensive

        ②Specificly introducing contained HPO algorithms

2.3. Related work

        ①Introduce other papers which test different HPO strategies

        ②Their testings are more comprehensive

2.4. Materials and methods

2.4.1. Dataset

        ①Dataset: Eterna, which includes mRNA mocule represented by sequence of nucleotides with structural features and targets

        ②With graph  constructed, nodes are nucleotides and edges are chemical bonds between them

        ③Number of molecules: 2400 in training set and 3634 in testing set

        ④Predicted metrics in training set are reactivity, deg_Mg_pH10, deg_Mg_50C’, deg_pH10’, deg_50C’, in tesing set are reactivity, deg_Mg_pH10, deg_Mg_50C’

        ⑤They further divided training set to 80% training and 20% validation

        ⑥Visualization of data:

2.4.2. Graph neural networks models

        ①Structure of models: graph stack and prediction stack

        ②Hyperparameters in the two parts:

（1）Graph Convolutional Network

（2）Graph Attention Network

2.4.3. Optimization algorithms

（1）Random search: randomly choose some combination. It will superior to grid search when only few parameters affect

（2）Bayesian Search: iteration global optimization method

（3）Hill climbing: updating hyperparameter vectors and adjast single hyperparameter each iteration 

（4）Simulated annealing: approximates global optimum

（5）Genetic algorithm: iteration and comibines the advantages in each generation

（6）Particle swarm optimization: similar to Greedy Algorithm?

（7）Artificial bee colony: neighbouring search 

2.5. Experiments, Results and Discussion

2.5.1. Experimental details

        ①Evaluation metric: Mean Column-wise Root Mean Squared Error (MCRMSE)

        ②Running time（时间不是次数）: 1h

        ③Meta hyperparameters of HPO: default

2.5.2. Results and discussion

        ①Performance table:

        ②Hyperparameter waves（这个颜色非常的明显！）:

        ③SA outperforms others

        ④⭐Mentioned common sets, namely their common excellent hyperparameter setting

2.6. Conclusion

        ①⭐They hope to contain more meta parameters as well

3. Reference

Vodilovska, V., Gievska, S. & Ivanoska, I. (2023) 'Hyperparameter Optimization of Graph Neural Networks for mRNA Degradation Prediction', 46th MIPRO ICT and Electronics Convention (MIPRO). doi:  10.23919/MIPRO57284.2023.10159737