Li Hongyi machine learning 2022-hw1

---

Preface

The article only serves as a record , Of course, there is reference to the ideas of the boss ヾ(•ω•`)o

---

Homework Introduction

---

kaggle link : link1
Google Colab link :link2
According to various states in the United States 4 Epidemic related data provided on the day , Forecast No 5 Day diagnostic rate

Standard for evaluation

Tips

---

Feature selection

---

This part can be first in jupyter or pycharm Medium python It will be more convenient to do it in the console
Read in the training set and test set , View the following numbers

Yes, of course , The training set has one more column than the test set , This column is the information that needs to be predicted , Find the column name of this column ‘tested_positive.4’, The corresponding column index is 117, That's the last column

Through the introduction, we know the former 37 The column data is the relevant information of the state , This part is represented by the unique heat code , Use here pandas Of corr Function analysis from 38 The data after the column is the same as ‘tested_positive.4’ The correlation between

Here I will make the correlation greater than 0.8 Select the features of , Pay attention to the last one ‘tested_positive.4’ You can't choose to go in , In the training stage, this is used to calculate loss Of

Get their index further

Yes select_feat Part of the code is modified

---

Different model architectures and optimizers

---

Because it is not big data , Increasing the depth of the model may lead to over fitting 、 Poor generalization and other problems ( I guess. ), Here we only change the width of the model

The optimizer consists of the original SGD Changed to Adam, Mu Shen said that it can reduce the impact of learning rate , There are also relevant explanations in Teacher Li Hongyi's subsequent courses

---

L2 regularization and try more parameters

L2 regularization Not really , Adjust the parameters slightly

---

Submit results

---

Finish the above steps and have a strong baseline There should be no problem , It may be better to adjust the parameters again ？ However, it is not recommended to adjust parameters repeatedly to achieve better results , This may lead to poor generalization of the model