# HTML Final
## 曾貴鴻
### data missing
[各種工具:連結](https://hcmy.gitbooks.io/ycimpute/content/)
(我們可以先都用這個tool ，用不同的填充方法，和不同的模型)
[sample_code](https://colab.research.google.com/drive/1LfD2ngz8IYKBdb1mUu0kDvinX_2lZAyd?usp=sharing)

feature:選出那些特徵使用(我將true,false轉成 1,0 ，且沒有使用任何文字資料)

第一次上傳2.3991 (by multiclass v1 knn=3)[連結](https://colab.research.google.com/drive/1umZPLBHCKu7kREm7ZPyZm-AVzDQtK6Dh#scrollTo=IywYnVSpIord)
第二次上傳2.38 (bymulticlass v2 knn=10)
第三次上傳2.37 (by multiclass v3 knn=10 , num of leaves =256)

### randomforest
classifier 1.8663366336633664
squerror 1.715492137449039
absolute 1.7958648806057076
squerror_weight1 1.7160745486313338
absolute_weight1 1.7891671520093186
squerror_weight2 1.7125800815375656
absolute_weight2 1.7900407687827606

next:加 Ein loss，增加樹木
(by max)
classifier best tree= 100 oob= 1.8829353523587653
classifier best tree= 150 oob= 1.8262085032032616
classifier best tree= 200 oob= 1.8217821782178218
classifier best tree= 250 oob= 1.8196854979615609
classifier best tree= 300 oob= 1.8082119976703552

(by median)
classifier best tree= 100 oob= 1.647932440302854
classifier best tree= 150 oob= 1.6314502038439138
classifier best tree= 200 oob= 1.6160745486313337

分析Ein error
ans= 0 average= 1.8990936555891238
[419, 376, 339, 235, 128, 91, 54, 11, 2, 0]
ans= 1 average= 3.0939597315436242
[64, 271, 370, 338, 227, 163, 133, 59, 14, 0]
ans= 2 average= 3.6466346153846154
[10, 122, 359, 352, 317, 247, 154, 79, 24, 0]
ans= 3 average= 4.155820895522388
[9, 57, 210, 375, 322, 318, 232, 128, 23, 1]
ans= 4 average= 4.563758389261745
[6, 37, 117, 234, 403, 378, 283, 142, 38, 1]
ans= 5 average= 4.853701527614571
[5, 20, 97, 202, 318, 484, 327, 206, 42, 1]
ans= 6 average= 5.348060344827586
[3, 9, 52, 148, 259, 435, 577, 291, 79, 3]
ans= 7 average= 5.76341730558598
[0, 9, 32, 79, 195, 371, 528, 493, 111, 8]
ans= 8 average= 6.27258064516129
[0, 2, 20, 50, 117, 284, 470, 578, 330, 9]
ans= 9 average= 6.887545344619105
[1, 3, 14, 14, 55, 123, 324, 559, 422, 139]

squerror best tree= 100 oob= 1.6960978450786255
squerror best tree= 150 oob= 1.6868957483983693
squerror best tree= 200 oob= 1.6857309260337798
squerror best tree= 200 oob= 1.6868375072801398
squerror best tree= 210 oob= 1.6863133372160746
squerror best tree= 220 oob= 1.6827606290040769

abs best tree= 100 oob= 1.7158998252766453
abs best tree= 150 oob= 1.7072801397786836
abs best tree= 200 oob= 1.7026790914385557
abs best tree= 250 oob= 1.7022714036109494

(by median with knn)
classifier best tree= 100 oob= 1.55713453698311
classifier best tree= 150 oob= 1.5395457192778101
classifier best tree= 200 oob= 1.5314502038439137
classifier best tree= 250 oob= 1.5263249854397205
classifier best tree= 300 oob= 1.519336051252184
classifier best tree= 350 oob= 1.5189866045428073
classifier best tree= 400 oob= 1.5152009318578916

## 蔡昀叡

[missing data填中位數 5foldCV + L1reg + huberloss+ onelayer 16->4->1model ###(error)2.48 ##validation loss 1.99](https://drive.google.com/file/d/1o0nClYNFth4P4aECvjaRMHZWdlwyB8oE/view?usp=sharing)
[跟上面一樣但model改成4層 16->64->16->16->1 ###(error) ##validation loss 1.96](https://drive.google.com/file/d/1P1gPs3EDGjkzHuhef7c04bYiCb91YhSh/view?usp=sharing)
[跟上面一樣但用KNN =4 ####error ##valid loss 1.938](https://drive.google.com/file/d/1lLDvoCAqmQCD4VVwQZ1SIpySDDc2-cLL/view?usp=sharing) [code](https://github.com/b08202011/Mlproject/blob/742a459b8a7e5211dc55d88f65f89e805a0c49cc/Knn%3D4%205foldCV%20%2B%20L1reg%20%2B%20%20huberloss%2B%20%204layer%2016-%3E64-%3E16-%3E16-%3E1model.ipynb)

valid loss用softL1loss加0.25是L1 loss
[先normalize softmax 算期望值 softL1loss 沒regulizer lr0.01 ###valid loss 1.233](https://drive.google.com/file/d/1Yhvh0hTAWcT8dl9K-i-Js_O_OiUjLG2s/view?usp=sharing)

新的loss 加了一個contraint 跟label距離超過2.5就會有 一個tube loss penalty
[multilayer penalty0.5 ##error2.11](https://drive.google.com/file/d/16nXELS6Qpw6FCR9P9WFgxPLCtRroBQPe/view?usp=sharing)
[maxpoollayer ##error2.11](https://drive.google.com/file/d/1uakHPeIzSxWmA0-OXOqrM0cz0Wlnat6R/view?usp=sharing)
[maxpoollayer penalty0.5 ##error2.18](https://drive.google.com/file/d/1XpFPihTzeH9mdy3zYZjyE4wc5oxK3QV-/view?usp=sharing)

## 陳毅安