# ml2019fall-hw1 Linear Regression

## Data Preprocessing

1. str.extrace('(Regular Expression)')
2. Concatanate

## Feature Selection

1. Covariance Matrix
2. L1 Norm Regularization => Implies Feature Selection => Sparse
* Seaborn heat map => weight $\approx$ 0.0 => useless feature will hurt the accuracy.
* 
3. L2 Norm Regularization => A Feasiable Function Set

+ [Feature Selection](https://taweihuang.hpd.io/2016/09/12/讀者提問：多元迴歸分析的變數選擇/)

## Improve

### In order to solve the overfitting problem

1. Regularization
* Elastic Network
3. Data augmentation
4. Early Stopping
5. Loss function
6. adam arguments

## Debug

1. sns heat map - cov matrix, weight
2. train loss v.s. val loss
3. Find some background knowledge (ex: 統計學、R 語言、PM 2.5、Github...)
4. Trial and Error

## Score

1. Public - Rank 45 - Score: 5.58173
2. Private - Rank 2 - Score: 5.24210