Bosch Production Line Performance, NO.74/top 6%

Bosch Production Line Performance e-mail : [email protected]

---

: 4000 Fitted model 50 top 6% 50

: R code parallel package mclapply linux mclapply sapply

ps :

---

1.

---/90%NA

NA AIC/BIC/lasso missing value XGB importance kernel Daniel FG code --- Feature Engineering

2.

Bosch Production Line Performance ( )

4000

Response	0	1
	1176868	6879

rate of Response 1 = 0.0058

2.1

Kaggle :

data	size	n ()	p ()	R ram
train_numeric	2.1GB	100	970	8.5 gb
train_date	2.9GB	100	1157	10.2 gb
train_categorical	2.7GB	100	2141
test_numeric	2.1GB	100	969
test_date	2.9GB	100	1157
test_categorical	2.7GB	100	2141

:

Response	, 0 : , 1 :
Id
Lx_Sx_Fx	L : lineS : stationF : feature number

L3_S36_F3939 3 36 3939 numeric date categorical

train_numeric :

Id	L0_S0_F0	L0_S0_F2	L0_S0_F4	L0_S0_F6	Response
11	-0.055	-0.086	0.294	0.330	0
13	0.003	0.019	0.294	0.312	0
14	NA	NA	NA	NA	0
16	NA	NA	NA	NA	0
18	-0.016	-0.041	-0.179	-0.179	0

train_date

Id	L0_S0_F0	L0_S0_F2	L0_S0_F4	L0_S0_F6
11	602.64	602.64	602.64	602.64
13	1331.66	1331.66	1331.66	1331.66
14	NA	NA	NA	NA
16	NA	NA	NA	NA
18	517.64	517.64	517.64	517.64

evaluation MCC

3.

3.1 feature engineering 1 ( 1 )

date data

feature
first
min
last
max
class.amount
na.amount	na

L0L1L2L3 ex : all_first, L0_first, L1_first, L2_first, L3_first data 100 feature feature engineering 1 kaggle rank 50% feature engineering 2

3.2 feature engineering 2 ( 2 )

feature		code
next
prev
total	total	next+prev
same.time		total>0
order.same.time		(cumsum(prev)+1) * same.time
group
group.length		table(group)
cost.time		max-min
prev.cost.time		cost.time-c(NA,cost.time[1:length(cost.time)-1])
next.cost.time		cost.time-c(cost.time[2:length(cost.time)],NA)
prev.na.amount	na	na.amount--c(NA,na.amount[1:length(na.amount)-1])
next.na.amount	na	na.amount--c(na.amount[2:length(na.amount)],NA)
prev.target		c(NA,target[1:(nrow(target)-1)])
next.target		c(target[2:nrow(target)],NA)

L0~L3

3.3

train_numeric

100 100

100 100 0.058 0.045 ID L3_S32_F3850

train_numeric :

-	res1.per	var.name
1	0.0451	L3_S32_F3850
2	0.0093	L1_S24_F1768
3	0.0093	L1_S24_F1763
.	...	...
.	...	...
968	0.0003	L1_S25_F2512

3.4 feature selection

feature engineering 1feature engineering 2 450 XGBoost xgb.cv bset nrounds bset nrounds

xgb.importance 50 50 fitted model feature

amount of var = 450

		pred
		0	1
real	0	1176353	515
	1	4216	2663

MCC = 0.568

amount of var = 50

		pred
		0	1
real	0	1176304	564
	1	4360	2519

MCC = 0.545

( 0 )( 1 )( )

xgb.cv bset nrounds nrounds xgb.cv nrounds

nrounds	train-rmse	test-rmse
11	0.168337	0.168881
21	0.083219	0.085046
31	0.064824	0.067989
41	0.061830	0.065582
51	0.061191	0.065279****
61	0.060756	0.065227
71	0.060327	0.065229

Best iteration : 67, train-rmse:0.060464 test-rmse:0.065220

nround = 50 test-rmse train & test model

3.5 other

imbalance evaluation --- MCC XGBoost evaluation

1. XGBoost rmse MCC
2. imbalance target
   0.25 0.25 1 0.25 0
   0 0

4. Fitted model

feature engineering 1feature engineering 2 train_numeric xgb.imporance 50 feature feature XGBoost evaluation MCC rmse imbalance rate = 0.25 imblance

Fitted model top 6% rank MCC 2 ( 0.18 -> 0.46 ) feature

50 feature

train_numeric

L3_S32_F3850**		L1_S24_F1723**		L3_S33_F3859**		L3_S33_F3855		L1_S24_F1846
L3_S33_F3865		L1_S24_F1632		L3_S33_F3857		L3_S38_F3956		L1_S24_F1498
L1_S24_F1604		L3_S41_F4014		L1_S24_F1695		L3_S38_F3952		L3_S33_F3873
L1_S24_F1844		L3_S38_F3960		L2_S26_F3036		L2_S26_F3040		L2_S26_F3047
L2_S26_F3073		L1_S24_F1672		L1_S24_F1609		L1_S24_F1685

feature of train_date

all_next**		next.cost.time**		next.traget**		L0_first**
all_prev**		prev.traget**		group.amount**		next.na.amount
all_na.amount		L3_first		total		cost.time
L2_first		L3_na.amount		prev.cost.time		group
L3_last		prev.na.amount		L3_min		L0_min
all_first		all_class.amount		L1_first		L3_max
order.same.time		L0_last

feature plot :

ML 1. 2. ( ) 3. modeltype 1 error vs typr 2 error 4. datamodel 100 datadatavsML DATA/ 5. DL/ensemble 6.

Reference

Bosch Production Line Performance. ( 2016 )

Daniel FG. ( 2016 )