Data ready
training_dataset.csv 를 분석을 위해 R session 으로 읽어들임
d <- paste0(data_path, "preprocessed/training_dataset.csv") %>% fread %>% tbl_dfMinor data preprocessing
분석의 편의를 위해 소소한 전처리를 아래와 같이 수행
date와time을 결합한 후 날짜시각 클레스로 관리하기 위하여dt필드로 변형dt로 날짜시각 클레스 파싱이 불가한 케이스의 제거1
- 요일 및 월, 일, 시각 변수를 파생시킴 :
wday,month,day,hour - 게절변수를 파생시켜
season필드 추가 swell을 numeric 에서 명목화
d <- d %>%
mutate(dt = paste(date, "_", time) %>% ymd_h,
month = lubridate::month(dt), wday = lubridate::wday(dt, label = T), day = lubridate::day(dt), hour = lubridate::hour(dt),
swell = as.factor(swell),
season = case_when(lubridate::month(dt) %in% c(3, 4, 5) ~ "spring",
lubridate::month(dt) %in% c(6, 7, 8, 9) ~ "summer",
lubridate::month(dt) %in% c(10, 11) ~ "autumn",
lubridate::month(dt) %in% c(12, 1, 2) ~ "winter") %>% factor(levels = c("spring", "summer", "autumn", "winter"))) %>%
filter(!is.na(dt)) %>%
select(-date, -time) %>%
select(dt, everything())EDA
Summary
필드별 값 현황 확인
summary(d)## dt swell wind_speed win_direction
## Min. :2014-01-04 15:00:00 0:4882 Min. : 0.000 Min. : 0.0
## 1st Qu.:2014-10-21 01:45:00 1:2578 1st Qu.: 4.400 1st Qu.: 44.0
## Median :2015-10-11 22:30:00 Median : 7.000 Median :131.0
## Mean :2015-10-25 02:01:06 Mean : 7.291 Mean :155.3
## 3rd Qu.:2016-10-08 10:15:00 3rd Qu.: 9.800 3rd Qu.:276.0
## Max. :2017-12-28 06:00:00 Max. :27.300 Max. :360.0
## NA's :461 NA's :401
## GUST air_pressure humidity temperature
## Min. : 0.000 Min. : 985.2 Min. :22.00 Min. :-5.70
## 1st Qu.: 5.800 1st Qu.:1008.5 1st Qu.:62.00 1st Qu.: 7.90
## Median : 9.100 Median :1015.0 Median :77.00 Median :15.20
## Mean : 9.406 Mean :1015.0 Mean :74.14 Mean :14.27
## 3rd Qu.: 12.500 3rd Qu.:1021.9 3rd Qu.:87.00 3rd Qu.:20.80
## Max. :110.000 Max. :1035.1 Max. :99.00 Max. :28.80
## NA's :456 NA's :136 NA's :178 NA's :149
## water_temperature max_wave_height mean_wave_height avg_wave_height
## Min. :10.00 Min. : 0.200 Min. :0.100 Min. :0.100
## 1st Qu.:13.40 1st Qu.: 1.500 1st Qu.:0.900 1st Qu.:0.700
## Median :18.20 Median : 2.600 Median :1.700 Median :1.200
## Mean :17.97 Mean : 2.833 Mean :1.793 Mean :1.272
## 3rd Qu.:22.30 3rd Qu.: 3.800 3rd Qu.:2.400 3rd Qu.:1.700
## Max. :29.10 Max. :11.700 Max. :7.700 Max. :5.500
## NA's :81 NA's :76 NA's :86 NA's :89
## wave_accurance wave_direction month wday
## Min. : 2.000 Min. : 0.0 Min. : 1.000 Sun: 956
## 1st Qu.: 5.300 1st Qu.: 79.0 1st Qu.: 3.000 Mon:1153
## Median : 7.100 Median :182.0 Median : 7.000 Tue:1117
## Mean : 7.054 Mean :176.5 Mean : 6.519 Wed:1190
## 3rd Qu.: 9.100 3rd Qu.:258.0 3rd Qu.:10.000 Thu:1083
## Max. :12.800 Max. :360.0 Max. :12.000 Fri: 962
## NA's :130 NA's :75 Sat: 999
## day hour season
## Min. : 1.00 Min. : 0.00 spring:1454
## 1st Qu.: 8.00 1st Qu.: 5.00 summer:2549
## Median :14.00 Median :11.00 autumn:1432
## Mean :14.97 Mean :11.05 winter:2025
## 3rd Qu.:22.00 3rd Qu.:17.00
## Max. :31.00 Max. :23.00
## 타겟변수인 swell 을 제외한 설명변수 중
연속형 변수들을의 시계열 도표를 그려봄
pd1_1 <- d %>%
select(-month, -wday, -day, -hour, -season) %>%
gather(class, value, -dt, -swell)
p1_1 <- pd1_1 %>%
ggplot(aes(dt, value, color = swell)) +
geom_step(stat = "identity") +
facet_wrap(~ class, ncol = 2, scales = "free")
p1_1 +
ggtitle("Time series plot by swell, class")
마찬가지로 시간별 분포는 어떠한지를 확인하기 위해 Boxplot 을 그려봄
pd1_2 <- pd1_1 %>%
mutate(hour = hour(dt) %>% as.factor)
p1_2 <- pd1_2 %>%
ggplot(aes(hour, value, color = swell)) +
geom_boxplot(outlier.alpha = 0.1) +
facet_wrap(~ class, ncol = 2, scales = "free")
p1_2 +
ggtitle("Boxplot group by hour by swell, class") +
labs(x = "hour")
마찬가지로 계절별 분포는 어떠한지 확인하기 위해 Boxplot 을 그려봄
pd1_3 <- d %>%
select(-month, -wday, -day, -hour, -dt) %>%
gather(class, value, -swell, -season)
p1_3 <- pd1_3 %>%
ggplot(aes(season, value, color = swell)) +
geom_boxplot(outlier.alpha = 0.1) +
facet_wrap(~ class, ncol = 2, scales = "free")
p1_3 +
ggtitle("Boxplot group by season by swell, class")
Boxplot 을 통한 swell 별 그리고 시간별, 계절별 패턴을 확인해 보면 패턴차이가 어느정도 존재하고, 이 패턴차를 이용하여 머신이 swell 을 구분하는 좋은 재료로 사용할 여지가 있어보임
독립변수간 상관관계를 확인하기 위해 상관행렬도를 그려봄
pd1_3 <- d %>%
select(-dt, -swell, -win_direction, -wave_direction, -month, -wday, -day, -hour, -season) %>%
cor(use = "pairwise.complete.obs")
corrplot(pd1_3, order = "AOE", method = "square")
corrplot(pd1_3, order = "AOE", type = "lower", method = "number", add = TRUE, diag = FALSE, tl.pos = "n", cl.pos = "n")
상관행렬도를 볼 때 상식적이지만 최대파고높이와 평균파고높이간에는 강력한 양의 상관관계가 있으며 두 변수는 거의 동일한 패턴이라고 볼 수 있다.
마찬가지로 풍속과 GUST 간에도 상당히 높은 양의 상관관계가 있는데 이런 동일패턴의 독립변수가 있는것이 얼추 확인 가능하고 이들은 머신러닝 학습시 중복된 정보로 치부될 가능성이 높아보임 2
Simple machine learning
간단한 머신러닝을 통해 학습 및 예측을 수행해 봄
Partition
학습 및 테스트 셋 분류
fullTrainset <- d %>%
select(-dt) %>%
na.omit
index <- createDataPartition(fullTrainset$swell, p = .7, list = F)
train <- fullTrainset[index, ]
test <- fullTrainset[-index, ]Training
학습모델은 mlMethods 를 통해 여러가지 알고리즘 모형을 적용하여 다양한 모델을 만들어 보고 상호 비교를 수행
하이퍼파라미터 최적 선정을 위한 validation 정책은 “5fold Cross Validation” 정책을 모든 알고리즘에 대해 통일하여 적용시킴
fitControl <- trainControl(method = "cv", number = 5, allowParallel = TRUE)
mlMethods <- c("simpls", "glm", "qda", "rocc", "knn", "rpart", "wsrf")
models <- mlMethods %>%
lapply(function(x) train(swell ~ ., data = train, method = x, trControl = fitControl))Evaluate
evaluate <- function(model, testset, class, ylim = c(1, testset %>% pull(class) %>% max)){
stopifnot(is.character(class))
real = pull(testset, class)
pred = predict(model, newdata = testset) %>% unlist
cm <- confusionMatrix(pred, real, positive = "1")
totalScore <- as.matrix(cm)[1, 1] + as.matrix(cm)[2, 2]*2 - as.matrix(cm)[2, 1] - as.matrix(cm)[1, 2]*2
paste0("<< Method = ", model$method,
" / Process time = ", model$times$everything[3] %>% round(3),
" / Total score = ", totalScore, " >>\n") %>%
cat
print(cm)
}혼동행렬 및 테스트셋 정답지를 이용한 점수계산(Total scroe)3 결과
for(i in mlMethods %>% length %>% seq) evaluate(models[[i]], test, "swell")## << Method = simpls / Process time = 1.143 / Total score = 1611 >>
## Confusion Matrix and Statistics
##
## Reference
## Prediction 0 1
## 0 1199 204
## 1 158 489
##
## Accuracy : 0.8234
## 95% CI : (0.8062, 0.8397)
## No Information Rate : 0.662
## P-Value [Acc > NIR] : < 2e-16
##
## Kappa : 0.5989
## Mcnemar's Test P-Value : 0.01802
##
## Sensitivity : 0.7056
## Specificity : 0.8836
## Pos Pred Value : 0.7558
## Neg Pred Value : 0.8546
## Prevalence : 0.3380
## Detection Rate : 0.2385
## Detection Prevalence : 0.3156
## Balanced Accuracy : 0.7946
##
## 'Positive' Class : 1
##
## << Method = glm / Process time = 0.982 / Total score = 1991 >>
## Confusion Matrix and Statistics
##
## Reference
## Prediction 0 1
## 0 1239 129
## 1 118 564
##
## Accuracy : 0.8795
## 95% CI : (0.8646, 0.8933)
## No Information Rate : 0.662
## P-Value [Acc > NIR] : <2e-16
##
## Kappa : 0.7297
## Mcnemar's Test P-Value : 0.5246
##
## Sensitivity : 0.8139
## Specificity : 0.9130
## Pos Pred Value : 0.8270
## Neg Pred Value : 0.9057
## Prevalence : 0.3380
## Detection Rate : 0.2751
## Detection Prevalence : 0.3327
## Balanced Accuracy : 0.8634
##
## 'Positive' Class : 1
##
## << Method = qda / Process time = 0.54 / Total score = 1951 >>
## Confusion Matrix and Statistics
##
## Reference
## Prediction 0 1
## 0 1209 124
## 1 148 569
##
## Accuracy : 0.8673
## 95% CI : (0.8519, 0.8817)
## No Information Rate : 0.662
## P-Value [Acc > NIR] : <2e-16
##
## Kappa : 0.706
## Mcnemar's Test P-Value : 0.1631
##
## Sensitivity : 0.8211
## Specificity : 0.8909
## Pos Pred Value : 0.7936
## Neg Pred Value : 0.9070
## Prevalence : 0.3380
## Detection Rate : 0.2776
## Detection Prevalence : 0.3498
## Balanced Accuracy : 0.8560
##
## 'Positive' Class : 1
##
## << Method = rocc / Process time = 1.796 / Total score = 1717 >>
## Confusion Matrix and Statistics
##
## Reference
## Prediction 0 1
## 0 1196 176
## 1 161 517
##
## Accuracy : 0.8356
## 95% CI : (0.8188, 0.8514)
## No Information Rate : 0.662
## P-Value [Acc > NIR] : <2e-16
##
## Kappa : 0.6307
## Mcnemar's Test P-Value : 0.4457
##
## Sensitivity : 0.7460
## Specificity : 0.8814
## Pos Pred Value : 0.7625
## Neg Pred Value : 0.8717
## Prevalence : 0.3380
## Detection Rate : 0.2522
## Detection Prevalence : 0.3307
## Balanced Accuracy : 0.8137
##
## 'Positive' Class : 1
##
## << Method = knn / Process time = 1.683 / Total score = 1969 >>
## Confusion Matrix and Statistics
##
## Reference
## Prediction 0 1
## 0 1232 131
## 1 125 562
##
## Accuracy : 0.8751
## 95% CI : (0.86, 0.8891)
## No Information Rate : 0.662
## P-Value [Acc > NIR] : <2e-16
##
## Kappa : 0.7204
## Mcnemar's Test P-Value : 0.7547
##
## Sensitivity : 0.8110
## Specificity : 0.9079
## Pos Pred Value : 0.8180
## Neg Pred Value : 0.9039
## Prevalence : 0.3380
## Detection Rate : 0.2741
## Detection Prevalence : 0.3351
## Balanced Accuracy : 0.8594
##
## 'Positive' Class : 1
##
## << Method = rpart / Process time = 0.904 / Total score = 1947 >>
## Confusion Matrix and Statistics
##
## Reference
## Prediction 0 1
## 0 1207 124
## 1 150 569
##
## Accuracy : 0.8663
## 95% CI : (0.8508, 0.8808)
## No Information Rate : 0.662
## P-Value [Acc > NIR] : <2e-16
##
## Kappa : 0.7041
## Mcnemar's Test P-Value : 0.131
##
## Sensitivity : 0.8211
## Specificity : 0.8895
## Pos Pred Value : 0.7914
## Neg Pred Value : 0.9068
## Prevalence : 0.3380
## Detection Rate : 0.2776
## Detection Prevalence : 0.3507
## Balanced Accuracy : 0.8553
##
## 'Positive' Class : 1
##
## << Method = wsrf / Process time = 44.04 / Total score = 2451 >>
## Confusion Matrix and Statistics
##
## Reference
## Prediction 0 1
## 0 1325 57
## 1 32 636
##
## Accuracy : 0.9566
## 95% CI : (0.9468, 0.965)
## No Information Rate : 0.662
## P-Value [Acc > NIR] : < 2e-16
##
## Kappa : 0.9021
## Mcnemar's Test P-Value : 0.01096
##
## Sensitivity : 0.9177
## Specificity : 0.9764
## Pos Pred Value : 0.9521
## Neg Pred Value : 0.9588
## Prevalence : 0.3380
## Detection Rate : 0.3102
## Detection Prevalence : 0.3259
## Balanced Accuracy : 0.9471
##
## 'Positive' Class : 1
## Case study
subset model : 시간적 설명변수를 모두 제거한 후 학습한 모델의 케이스
sub_models <- mlMethods %>%
lapply(function(x) train(swell ~ . - month - wday - day - hour - season, data = train, method = x, trControl = fitControl))
for(i in mlMethods %>% length %>% seq) evaluate(sub_models[[i]], test, "swell")## << Method = simpls / Process time = 0.604 / Total score = 1599 >>
## Confusion Matrix and Statistics
##
## Reference
## Prediction 0 1
## 0 1199 207
## 1 158 486
##
## Accuracy : 0.822
## 95% CI : (0.8047, 0.8383)
## No Information Rate : 0.662
## P-Value [Acc > NIR] : < 2e-16
##
## Kappa : 0.5952
## Mcnemar's Test P-Value : 0.01199
##
## Sensitivity : 0.7013
## Specificity : 0.8836
## Pos Pred Value : 0.7547
## Neg Pred Value : 0.8528
## Prevalence : 0.3380
## Detection Rate : 0.2371
## Detection Prevalence : 0.3141
## Balanced Accuracy : 0.7924
##
## 'Positive' Class : 1
##
## << Method = glm / Process time = 0.605 / Total score = 1977 >>
## Confusion Matrix and Statistics
##
## Reference
## Prediction 0 1
## 0 1240 133
## 1 117 560
##
## Accuracy : 0.878
## 95% CI : (0.8631, 0.8919)
## No Information Rate : 0.662
## P-Value [Acc > NIR] : <2e-16
##
## Kappa : 0.726
## Mcnemar's Test P-Value : 0.3428
##
## Sensitivity : 0.8081
## Specificity : 0.9138
## Pos Pred Value : 0.8272
## Neg Pred Value : 0.9031
## Prevalence : 0.3380
## Detection Rate : 0.2732
## Detection Prevalence : 0.3302
## Balanced Accuracy : 0.8609
##
## 'Positive' Class : 1
##
## << Method = qda / Process time = 0.489 / Total score = 1971 >>
## Confusion Matrix and Statistics
##
## Reference
## Prediction 0 1
## 0 1211 120
## 1 146 573
##
## Accuracy : 0.8702
## 95% CI : (0.8549, 0.8845)
## No Information Rate : 0.662
## P-Value [Acc > NIR] : <2e-16
##
## Kappa : 0.7127
## Mcnemar's Test P-Value : 0.1253
##
## Sensitivity : 0.8268
## Specificity : 0.8924
## Pos Pred Value : 0.7969
## Neg Pred Value : 0.9098
## Prevalence : 0.3380
## Detection Rate : 0.2795
## Detection Prevalence : 0.3507
## Balanced Accuracy : 0.8596
##
## 'Positive' Class : 1
##
## << Method = rocc / Process time = 1.218 / Total score = 1601 >>
## Confusion Matrix and Statistics
##
## Reference
## Prediction 0 1
## 0 1246 230
## 1 111 463
##
## Accuracy : 0.8337
## 95% CI : (0.8168, 0.8495)
## No Information Rate : 0.662
## P-Value [Acc > NIR] : < 2.2e-16
##
## Kappa : 0.612
## Mcnemar's Test P-Value : 1.658e-10
##
## Sensitivity : 0.6681
## Specificity : 0.9182
## Pos Pred Value : 0.8066
## Neg Pred Value : 0.8442
## Prevalence : 0.3380
## Detection Rate : 0.2259
## Detection Prevalence : 0.2800
## Balanced Accuracy : 0.7932
##
## 'Positive' Class : 1
##
## << Method = knn / Process time = 1.042 / Total score = 1919 >>
## Confusion Matrix and Statistics
##
## Reference
## Prediction 0 1
## 0 1233 144
## 1 124 549
##
## Accuracy : 0.8693
## 95% CI : (0.8539, 0.8836)
## No Information Rate : 0.662
## P-Value [Acc > NIR] : <2e-16
##
## Kappa : 0.7058
## Mcnemar's Test P-Value : 0.2458
##
## Sensitivity : 0.7922
## Specificity : 0.9086
## Pos Pred Value : 0.8158
## Neg Pred Value : 0.8954
## Prevalence : 0.3380
## Detection Rate : 0.2678
## Detection Prevalence : 0.3283
## Balanced Accuracy : 0.8504
##
## 'Positive' Class : 1
##
## << Method = rpart / Process time = 0.696 / Total score = 1947 >>
## Confusion Matrix and Statistics
##
## Reference
## Prediction 0 1
## 0 1207 124
## 1 150 569
##
## Accuracy : 0.8663
## 95% CI : (0.8508, 0.8808)
## No Information Rate : 0.662
## P-Value [Acc > NIR] : <2e-16
##
## Kappa : 0.7041
## Mcnemar's Test P-Value : 0.131
##
## Sensitivity : 0.8211
## Specificity : 0.8895
## Pos Pred Value : 0.7914
## Neg Pred Value : 0.9068
## Prevalence : 0.3380
## Detection Rate : 0.2776
## Detection Prevalence : 0.3507
## Balanced Accuracy : 0.8553
##
## 'Positive' Class : 1
##
## << Method = wsrf / Process time = 28.352 / Total score = 2403 >>
## Confusion Matrix and Statistics
##
## Reference
## Prediction 0 1
## 0 1317 65
## 1 40 628
##
## Accuracy : 0.9488
## 95% CI : (0.9383, 0.9579)
## No Information Rate : 0.662
## P-Value [Acc > NIR] : < 2e-16
##
## Kappa : 0.8845
## Mcnemar's Test P-Value : 0.01917
##
## Sensitivity : 0.9062
## Specificity : 0.9705
## Pos Pred Value : 0.9401
## Neg Pred Value : 0.9530
## Prevalence : 0.3380
## Detection Rate : 0.3063
## Detection Prevalence : 0.3259
## Balanced Accuracy : 0.9384
##
## 'Positive' Class : 1
## Subset model 과 Full model 과 비교해 볼 때
Total score 가 미세하지만 상대적으로 떨어지는 경향이 보이므로, 파생시킬 수 있는 시간적 설명변수를 최대한 고려해 보는게 유리할 것으로 판단됨
Advanced analysis
분산팽창지수(VIF) 체크
편의상 설명변수 중 연속형 변수간 다중공산성을 분산팽창지수를 통해 확인
forVif <- fullTrainset %>% select(-month, -day, -wday, -hour, -season) %>% mutate(swell = as.numeric(swell))
lm(swell ~ ., data = forVif) %>% vif## wind_speed win_direction GUST air_pressure
## 17.353791 1.264370 19.544251 2.751693
## humidity temperature water_temperature max_wave_height
## 2.116257 10.024319 5.898366 24.514069
## mean_wave_height avg_wave_height wave_accurance wave_direction
## 458.705780 453.460356 2.470627 1.041055lm(swell ~ . - mean_wave_height, data = forVif) %>% vif## wind_speed win_direction GUST air_pressure
## 17.343611 1.264325 19.526862 2.751671
## humidity temperature water_temperature max_wave_height
## 2.116257 10.021081 5.897988 24.059268
## avg_wave_height wave_accurance wave_direction
## 24.750127 2.469571 1.040679lm(swell ~ . - mean_wave_height - avg_wave_height, data = forVif) %>% vif## wind_speed win_direction GUST air_pressure
## 17.263232 1.258364 19.523719 2.747348
## humidity temperature water_temperature max_wave_height
## 2.116230 10.015590 5.897917 3.605369
## wave_accurance wave_direction
## 2.375875 1.040643lm(swell ~ . - mean_wave_height - avg_wave_height - GUST, data = forVif) %>% vif## wind_speed win_direction air_pressure humidity
## 2.044936 1.256764 2.692634 2.095751
## temperature water_temperature max_wave_height wave_accurance
## 9.466656 5.610249 3.407452 2.347677
## wave_direction
## 1.0309443번의 스텝을 통해 분산팽창지수가 10 이하인 변수만을 선별할 경우 mean_wave_height, avg_wave_height, GUST 변수가 탈락됨
따라서 본 3개의 변수를 제외시킨 Subset model 로 예측성능을 확인해 보면 아래와 같음
sub_models <- mlMethods %>%
lapply(function(x) train(swell ~ . - mean_wave_height - avg_wave_height - GUST, data = train, method = x, trControl = fitControl))
for(i in mlMethods %>% length %>% seq) evaluate(sub_models[[i]], test, "swell")## << Method = simpls / Process time = 0.645 / Total score = 1609 >>
## Confusion Matrix and Statistics
##
## Reference
## Prediction 0 1
## 0 1198 204
## 1 159 489
##
## Accuracy : 0.8229
## 95% CI : (0.8057, 0.8392)
## No Information Rate : 0.662
## P-Value [Acc > NIR] : < 2e-16
##
## Kappa : 0.598
## Mcnemar's Test P-Value : 0.02092
##
## Sensitivity : 0.7056
## Specificity : 0.8828
## Pos Pred Value : 0.7546
## Neg Pred Value : 0.8545
## Prevalence : 0.3380
## Detection Rate : 0.2385
## Detection Prevalence : 0.3161
## Balanced Accuracy : 0.7942
##
## 'Positive' Class : 1
##
## << Method = glm / Process time = 0.69 / Total score = 1999 >>
## Confusion Matrix and Statistics
##
## Reference
## Prediction 0 1
## 0 1239 127
## 1 118 566
##
## Accuracy : 0.8805
## 95% CI : (0.8657, 0.8942)
## No Information Rate : 0.662
## P-Value [Acc > NIR] : <2e-16
##
## Kappa : 0.7321
## Mcnemar's Test P-Value : 0.6093
##
## Sensitivity : 0.8167
## Specificity : 0.9130
## Pos Pred Value : 0.8275
## Neg Pred Value : 0.9070
## Prevalence : 0.3380
## Detection Rate : 0.2761
## Detection Prevalence : 0.3337
## Balanced Accuracy : 0.8649
##
## 'Positive' Class : 1
##
## << Method = qda / Process time = 0.557 / Total score = 1939 >>
## Confusion Matrix and Statistics
##
## Reference
## Prediction 0 1
## 0 1209 127
## 1 148 566
##
## Accuracy : 0.8659
## 95% CI : (0.8503, 0.8803)
## No Information Rate : 0.662
## P-Value [Acc > NIR] : <2e-16
##
## Kappa : 0.7025
## Mcnemar's Test P-Value : 0.2278
##
## Sensitivity : 0.8167
## Specificity : 0.8909
## Pos Pred Value : 0.7927
## Neg Pred Value : 0.9049
## Prevalence : 0.3380
## Detection Rate : 0.2761
## Detection Prevalence : 0.3483
## Balanced Accuracy : 0.8538
##
## 'Positive' Class : 1
##
## << Method = rocc / Process time = 1.372 / Total score = 1717 >>
## Confusion Matrix and Statistics
##
## Reference
## Prediction 0 1
## 0 1196 176
## 1 161 517
##
## Accuracy : 0.8356
## 95% CI : (0.8188, 0.8514)
## No Information Rate : 0.662
## P-Value [Acc > NIR] : <2e-16
##
## Kappa : 0.6307
## Mcnemar's Test P-Value : 0.4457
##
## Sensitivity : 0.7460
## Specificity : 0.8814
## Pos Pred Value : 0.7625
## Neg Pred Value : 0.8717
## Prevalence : 0.3380
## Detection Rate : 0.2522
## Detection Prevalence : 0.3307
## Balanced Accuracy : 0.8137
##
## 'Positive' Class : 1
##
## << Method = knn / Process time = 1.445 / Total score = 1945 >>
## Confusion Matrix and Statistics
##
## Reference
## Prediction 0 1
## 0 1234 138
## 1 123 555
##
## Accuracy : 0.8727
## 95% CI : (0.8575, 0.8868)
## No Information Rate : 0.662
## P-Value [Acc > NIR] : <2e-16
##
## Kappa : 0.714
## Mcnemar's Test P-Value : 0.3862
##
## Sensitivity : 0.8009
## Specificity : 0.9094
## Pos Pred Value : 0.8186
## Neg Pred Value : 0.8994
## Prevalence : 0.3380
## Detection Rate : 0.2707
## Detection Prevalence : 0.3307
## Balanced Accuracy : 0.8551
##
## 'Positive' Class : 1
##
## << Method = rpart / Process time = 0.88 / Total score = 1947 >>
## Confusion Matrix and Statistics
##
## Reference
## Prediction 0 1
## 0 1207 124
## 1 150 569
##
## Accuracy : 0.8663
## 95% CI : (0.8508, 0.8808)
## No Information Rate : 0.662
## P-Value [Acc > NIR] : <2e-16
##
## Kappa : 0.7041
## Mcnemar's Test P-Value : 0.131
##
## Sensitivity : 0.8211
## Specificity : 0.8895
## Pos Pred Value : 0.7914
## Neg Pred Value : 0.9068
## Prevalence : 0.3380
## Detection Rate : 0.2776
## Detection Prevalence : 0.3507
## Balanced Accuracy : 0.8553
##
## 'Positive' Class : 1
##
## << Method = wsrf / Process time = 37.078 / Total score = 2445 >>
## Confusion Matrix and Statistics
##
## Reference
## Prediction 0 1
## 0 1326 59
## 1 31 634
##
## Accuracy : 0.9561
## 95% CI : (0.9463, 0.9646)
## No Information Rate : 0.662
## P-Value [Acc > NIR] : < 2.2e-16
##
## Kappa : 0.9009
## Mcnemar's Test P-Value : 0.004427
##
## Sensitivity : 0.9149
## Specificity : 0.9772
## Pos Pred Value : 0.9534
## Neg Pred Value : 0.9574
## Prevalence : 0.3380
## Detection Rate : 0.3093
## Detection Prevalence : 0.3244
## Balanced Accuracy : 0.9460
##
## 'Positive' Class : 1
## Fullset 에 비해 큰 차이가 없음
이 말은 역으로 생각하면 mean_wave_height, avg_wave_height, GUST 설명변수를 포함하든 그렇지 않든 예측성능에 큰 영향력이 없다는 것으로 볼 수 있으며,
따라서 모형의 복잡도를 낮추는 측면에서 패턴이 비슷한 3개의 설명변수를 학습셋에 고려하지 않는 편이 유리할 수 있음
Models Evaluation table
evaluationRes <- data.frame(Method = c("simpls", "glm", "qda", "rocc", "knn", "rpart", "wsrf"),
Method_Fullname = c("Partial Least Squares", "Generalized Linear Model", "Quadratic Discriminant Analysis", "ROC-Based Classifier", "k-Nearest Neighbors", "CART", "Weighted Subspace Random Forest"),
FullModel = c(1585, 1949, 1991, 1589, 1983, 1879, 2419),
SubsetModel_timeOmit = c(1575, 1963, 2061, 1589, 1947, 1879, 2351),
SubsetModel_afterVIF = c(1581, 1947, 1975, 1589, 1973, 1879, 2441))
evaluationRes %>%
datatable(options = list(pageLength = 30)) %>%
formatStyle("FullModel", background = styleColorBar(evaluationRes$FullModel, "steelblue"))