공공데이터를 활용한 특정지역 온도추정 경진대회 1위 수상작입니다. https://dacon.io 에서 더 많은 정보를 확인할 수 있습니다.

1. https://dacon.io
AI프렌즈 시즌1
공공데이터를 활용한
온도추정 경진대회
최상혁

2. 목차
1
2
3
STEP 1
STEP 2
STEP 3
https://dacon.io 2
EDA & 데이터 전처리
모델 구축 & 검증
결과 및 결언
EDA & 데이터 전처리
모델 구축 & 검증
결과 및 결언
• Single Model
• Multi Model
• Conclusion
• Suggestions
• Library & Loading
• Cleansing
• Feature Engineering
• Feature Selection

3. https://dacon.io 3
1. EDA & 데이터 전처리 : Library & Loading

4. https://dacon.io 4
1. EDA & 데이터 전처리 : Library & Loading

5. https://dacon.io 5
1. EDA & 데이터 전처리1. EDA & 데이터 전처리 : Cleansing

6. https://dacon.io 6
1. EDA & 데이터 전처리 : Feature Engineering
일일 누적 -> 1시간 누적

7. https://dacon.io 7
1. EDA & 데이터 전처리 : Feature Engineering
Before After
일사량
Time
일사량
Time
일사량 일사량
Y 평균
or Y18
Y 평균
or Y18
해 뜨기 전 해 진 후

8. https://dacon.io 8
1. EDA & 데이터 전처리 : Feature Selection
기온 vs mean of Ys / Y18
●파란색 : Train1 Data
●빨간색 : Train2 Data
일사량 vs mean of Ys / Y18
습도 vs mean of Ys / Y18
기온
일사량
습도

9. https://dacon.io 9
기온, 일사량, 습도 외
나머지 feature들은 모두 Drop!
1. EDA & 데이터 전처리 : Feature Selection

10. https://dacon.io 10
1. EDA & 데이터 전처리 : Feature Engineering
기온과 습도 각각에서 몇개의 변수들로 mean features를 생성
<- X00 제외
<- X12, X20 제외

11. https://dacon.io
1. EDA & 데이터 전처리 : Feature Engineering
그림자?
↓
Negative
Insolation
(- insolation)
11

12. https://dacon.io 12
1. EDA & 데이터 전처리 : Feature Selection
Noisy Feature를 찾아서 제거!

13. https://dacon.io 13
2. 모델 구축 & 검증 : Single Model
Y00
-
Y17
Y00
-
Y17
Predict
Predict
Fit
XTest :
Y00
-
Y17
XTrain 1 :
XTrain 2 : Y18Fit
Y18Predict
Evaluation :
Training score ,
Visualization,
Public score
Learning Algorithm :
Lasso (LassoCV)
= "Y들"
단순한 구조!!

14. https://dacon.io 14
2. 모델 구축 & 검증 : Single Model

15. https://dacon.io 15
2. 모델 구축 & 검증 : Single Model
X['time'] = X.index % 144
Day Time : [34 – 100] Night Time : [0 – 33] & [101 - 143]
test
Y18
Day index = 33 / Night index = 100

16. https://dacon.io 16
2. 모델 구축 & 검증 : Single Model

17. https://dacon.io 17
2. 모델 구축 & 검증 : Multi Model
Day index = 21 / Night index = 91 Day index = 21 / Night index = 93 Day index = 21 / Night index = 95
Day index = 43 / Night index = 103 Day index = 43 / Night index = 105 Day index = 43 / Night index = 107
·····
Y18
Day index : np.arange(21, 44, 2) <- 12개
Night index : np.arange(91, 108, 2) <- 9개
# of Sub models : 12 * 9 = 108
"Sub model"

18. https://dacon.io 18
2. 모델 구축 & 검증 : Single Model
Day index = 21 / Night index = 91
·····
Fit
Y00
-
Y17
X
Day Time : [15 – 85]
Predict
Y00
-
Y17
X
Train1 :
Train2 :
Fit
Y00
-
Y17
X
Day Time : [18 – 85]
Predict
Y00
-
Y17
X
Train1 :
Train2 :
Fit
Y00
-
Y17
X
Day Time : [21 – 85]
Predict
Y00
-
Y17
X
Train1 :
Train2 :
Fit
Y00
-
Y17
X
Day Time : [21 – 97]
Predict
Y00
-
Y17
X
Train1 :
Train2 :
Fit
Y00
-
Y17
X
Day Time : [24 – 97]
Predict
Y00
-
Y17
X
Train1 :
Train2 :
Fit
Y00
-
Y17
X
Day Time : [27 – 97]
Predict
Y00
-
Y17
X
Train1 :
Train2 :
Day index extended = [15, 18, 21, 24, 27] / Night index extended = [85, 88, 91, 94, 97]
np.meshgrid([15, 18, 21, 24, 27], [85, 88, 91, 94, 97]) <- 총 25개
"Sub-sub model"
↑ Day 예시
Night도 똑같음
Y00
-
Y17
# Sub-sub models : 5 * 5 = 25

19. https://dacon.io 19
2. 모델 구축 & 검증 : Multi Model
Day index = 21 / Night index = 91 Day index = 21 / Night index = 93 Day index = 21 / Night index = 95
Day index = 43 / Night index = 103 Day index = 43 / Night index = 105 Day index = 43 / Night index = 107
·····
Y18
× 25 × 25 × 25 × 25 × 25 × 25
× 25× 25× 25× 25× 25× 25
# Sub models : 12 * 9 = 108 108 (# Sub models) * 2 (day, night)
* 25 (# Sub-sub models) * 18 (Y00~Y17)
= 97200

20. https://dacon.io 20
2. 모델 구축 & 검증 : Multi Model
·
·
·
다시 Fit 할 필요 없이 바로 Predict 가능
일반적인 model
Train_X Train_YFit
Predict Test Y_pred
Predict2 Test2 Y_pred2

21. https://dacon.io 21
2. 모델 구축 & 검증 : Multi Model
Day index = 21 / Night index = 91 Day index = 21 / Night index = 93 Day index = 21 / Night index = 95
Day index = 43 / Night index = 103 Day index = 43 / Night index = 105 Day index = 43 / Night index = 107
·····
Y18
Test2 Test2 Test2 Test2 Test2 Test2
Test2Test2Test2Test2Test2Test2
× 25 × 25 × 25 × 25 × 25 × 25
× 25× 25× 25× 25× 25× 25
Test2다시 Training...?
108 (# Sub models) * 2 (day, night)
* 25 (# Sub-sub models) * 18 (Y00~Y17)
= 97200

22. https://dacon.io
2. 모델 구축 & 검증 : Multi Model
× 25
Y00
-
Y17
Y00
-
Y17
Predict
Predict
Fit
XTest :
Y00
-
Y17
XTrain 1 :
XTrain 2 : Y18Fit
Y18Predict
Y00
-
Y17
Predict
Fit
Y00
-
Y17
XTrain 1 :
XTrain 2 : Y18Fit
× 25
model1.coef_ Y00 Y01 Y02
X00 0.1 2.1 1.5
X01 2.7 3.4 5.1
X02 6.3 5.7 3.3
Ys_Intercept 15.0 13.3 16.8
22
model2.coef_ Y18
Y00 0.5
Y01 1.2
Y02 -1.3
Y18_Intercept 2.1
model1
X -> {Y00-Y17}
model2
{Y00-Y17} -> Y18
X -> {Y00-Y17} {Y00-Y17} -> Y18
결과값이 아닌
Coefficients의 평균!

23. https://dacon.io 23
2. 모델 구축 & 검증 : Multi Model
model1.coef_ Y00 Y01 Y02
X00 0.1 2.1 1.5
X01 2.7 3.4 5.1
X02 6.3 5.7 3.3
Ys_Intercept 15.0 13.3 16.8
model2.coef_ Y18
Y00 0.5
Y01 1.2
Y02 -1.3
Y18_Intercept 2.1
[model1 + model2]
X -> {Y00-Y17} -> Y18
×
0.5
×
1.2
×
-1.3
model1 + model2 Y18
X00 0.62
X01 -1.2
X02 5.7
Ys_Intercept 1.62
Y18_Intercept 2.1
model1.coef_ Y00 Y01 Y02
X00 0.05 2.52 -1.95
X01 1.35 4.08 -6.63
X02 3.15 6.84 -4.29
Ys_Intercept 7.5 15.96 -21.84
Coefficients of
Mapping
X -> Y18

24. https://dacon.io 24
2. 모델 구축 & 검증 : Multi Model
Coefficients of
Mapping
X -> Y18
Y00
-
Y17
Predict
Fit
Y00
-
Y17
XTrain 1 :
XTrain 2 : Y18Fit
× 25

25. 25
2. 모델 구축 & 검증 : Multi Model
Day Time : [34 – 100] Night Time : [0 – 33] & [101 - 143]
DAY
Coefs
NIGHT
Coefs
Time X00 X01 ··· neg_insol Ys_intercept Y18_intercept
0 1.3 0.0 ··· 0.0 3.8 1.2
1 1.3 0.0 ··· 0.0 3.8 1.2
··· 1.3 0.0 ··· 0.0 3.8 1.2
34 0.0 5.5 ··· -2.4 1.3 0.5
··· 0.0 5.5 ··· -2.4 1.3 0.5
100 0.0 5.5 ··· -2.4 1.3 0.5
··· 1.3 0.0 ··· 0.0 3.8 1.2
143 1.3 0.0 ··· 0.0 3.8 1.2
Day coefs
Night coefs
Night coefs
Coef_df :
Coefs_df
"Sub model"

26. https://dacon.io 26
2. 모델 구축 & 검증 : Multi Model
·····
Coefs_df
Day index = 21 / Night index = 93 Day index = 21 / Night index = 95Day index = 21 / Night index = 91
Day index = 43 / Night index = 103 Day index = 43 / Night index = 105 Day index = 43 / Night index = 107
# Sub models : 12 * 9 = 108

27. https://dacon.io 27
2. 모델 구축 & 검증 : Multi Model
time id X00 X01 ··· neg_insol Ys_intercept Y18_intercept
100 12340 26.1 45.0 ··· 1.1 1.0 1.0
101 12341 26.2 44.8 ··· 1.0 1.0 1.0
102 12342 26.4 44.2 ··· 0.8 1.0 1.0
Time X00 X01 ··· neg_insol Ys_intercept Y18_intercept
0 1.3 0.0 ··· 0.0 3.8 1.2
1 1.3 0.0 ··· 0.0 3.8 1.2
··· 1.3 0.0 ··· 0.0 3.8 1.2
34 0.3 7.8 ··· -0.6 2.5 1.1
··· 0.0 5.5 ··· -2.4 1.3 0.5
100 0.9 4.2 ··· -0.3 3.1 0.8
··· 1.3 0.0 ··· 0.0 3.8 1.2
143 1.3 0.0 ··· 0.0 3.8 1.2
Coefs_df :
Test:
(after preprocessing)
Dot Product!

28. https://dacon.io 28
3. 결과 및 결언 : Conclusion
Public score : 1.02845
Private score : 2.67974

29. https://dacon.io 29
3. 결과 및 결언 : Conclusion

30. https://dacon.io 30
3. 결과 및 결언 : Suggestion
맑은 날 vs 흐린 날을 구분할 수 있는
좀 더 정교한 feature를 도입했더라면...
다양한 알고리즘의 시도
+ Transfer Learning ?

31. THANK YOU
THANK YOU
https://dacon.io 31