## I'm so sorry about my English skill... I'm not good at English ## I'll not give you output. Because 'H2o' Package that useful DeepLeaning analysis tool ## must link each Computer's java programe and R programe. So I couldn't display output. ## Importing packages #install.packages("readr") #install.packages("h2o") #library(readr) #library(h2o) ## These libraries are my favorite R libraries. There are include many functions in h2o. # ### Running code #bat <- read_csv("C:\\Users\\LGPC\\Documents\\train_V2.csv") #bat <- as.data.frame(bat) #rank <- bat[, c(8, 16, 25)] ; rank[1:20 , ] #summary(rank[rank$killPoints > 0, ]$killPoints) #summary(rank[rank$rankPoints > 0, ]$rankPoints) #summary(rank[rank$winPoints > 0, ]$winPoints) #rank <- ifelse(rank$rankPoints > 0, rank$rankPoints, (rank$winPoints - 16)) #quantile(rank, probs = seq(0, 1, .1)) ; quantile(rank, probs = c(.005, .995)) # ## Points merge #bat$rank <- rank ; bat$killPoints = NULL ; bat$winPoints = NULL ; bat$rankPoints = NULL ## use .5% ~ 99.5% data. #bat <- bat[(bat$rank >= 1245) & (bat$rank <= 1837), ] # ## delete outlier (MatchDuration) #quantile(bat$matchDuration, probs = seq(0, .1, .005)) #bat <- bat[bat$matchDuration >= 1167, ] #ng <- bat$numGroups ; summary(ng) ; quantile(ng, probs = seq(0, .1, .005)) ## delete 'numGruops' under 16 #bat <- bat[bat$numGroups >= 16, ]; ng <- ng[ng >= 16] #bat$maxPlace = NULL ; bat$numGroups = NULL # ## k-Means clustering about match Type. biggest number of group is Solo data #set.seed(123) #ng_km <- kmeans(ng, center = 3, iter.max = 7000) ; ng_km # duo:1 , solo:2 , squad:3 #MT <- bat$matchType # #bat$per <- bat$winPlacePerc ; bat$winPlacePerc = NULL ; bat$matchType = NULL # ## delete unuseful feature #full.fit <- lm(per ~ ., data = bat) ; summary(full.fit) ; vif(full.fit) #bat$kills = NULL #full.fit <- lm(per ~ ., data = bat) ; summary(full.fit) ; vif(full.fit) #bat$headshotKills = NULL ; bat$roadKills = NULL ; bat$vehicleDestroys = NULL ; # ## K-Means clustering about matchDuration. We can splite matchDuration into BattleRoyal and MiniRoyal. #md <- bat$matchDuration #set.seed(123) ; kmd <- kmeans(md, centers = 2, iter.max = 7000) ; bat$matchDuration <- NULL ## bat : 1 , mini : 2 #bat$bat_or_mini <- kmd$cluster # #solo$assists = 0 ; solo$DBNOs = 0 ; solo$revives = 0 ; solo$teamKills = 0 # #bom <- solo$bat_or_mini ; solo$bat_or_mini = NULL #bat <- solo[bom == 1, ] ; mini <- solo[bom == 2, ] # ## Use H2o Package. #solo <- as.h2o(solo) ; bat <- as.h2o(bat) ; mini <- as.h2o(mini) #rand <- h2o.runif(solo, seed = 123) #solotr <- solo[rand <= .7, ] ; solote <- solo[rand > .7, ] #rand <- h2o.runif(bat, seed = 123) #battr <- bat[rand <= .7, ] ; batte <- bat[rand > .7, ] #rand <- h2o.runif(mini, seed = 123) #minitr <- mini[rand <= .7, ] ; minite <- mini[rand > .7, ] # ### Use RandomForest Model #b.rf.model <- h2o.randomForest(x = 1:15, # y = 16, # training_frame = battr, # validation_frame = batte, # model_id = "b.rf.model", # distribution = "gaussian", # ntree = 175) #m.rf.model <- h2o.randomForest(x = 1:15, # y = 16, # training_frame = minitr, # validation_frame = minite, # model_id = "m.rf.model", # distribution = "gaussian", # ntree = 175) #a1 <- h2o.predict(b.rf.model, batte) #a1 <- ifelse(a1 > 1, 1, a1) #a1 <- ifelse(a1 < 0, 0, a1) #b.rf.res <- a1 - batte$per ## Solo BattleRoyal's MSE : .00652 #mean(b.rf.res^2) #a2 <- h2o.predict(m.rf.model, minite) #a2 <- ifelse(a2 > 1, 1, a2) #a2 <- ifelse(a2 < 0, 0, a2) #m.rf.res <- a2 - minite$per ## Solo MiniRoyal's MSE : .00603 #mean(m.rf.res^2) # ### Use generalized boosted regression Model #b.gbm.model <- h2o.gbm(x = 1:15, # y = 16, # training_frame = battr, # validation_frame = batte, # nfolds = 5, # fold_assignment = "Random", # ntrees = 175, # seed = 123) #m.gbm.model <- h2o.gbm(x = 1:15, # y = 16, # training_frame = minitr, # validation_frame = minite, # fold_assignment ="Random", # distribution = "gaussian", # ntrees = 175, # nfolds = 5, # seed = 123) #a1 <- h2o.predict(b.gbm.model, batte) #a1 <- ifelse(a1 > 1, 1, a1) #a1 <- ifelse(a1 < 0, 0, a1) #b.gbm.res <- a1 - batte$per ## .004557565 #mean(b.gbm.res^2) #a2 <- h2o.predict(m.gbm.model, minite) #a2 <- ifelse(a2 > 1, 1, a2) #a2 <- ifelse(a2 < 0, 0, a2) #m.gbm.res <- a2 - minite$per ## .003985434 #mean(m.gbm.res^2) # ### So I decided Use H2o's GBM Modeling ### I made all about model solo, duo, squad, flare, crash. ### But flare and crash type's MSE are low when I use RandomForest Model. #solobat.model <- h2o.gbm(x = 1:15, # y = 16, # training_frame = bat, # nfolds = 5, # fold_assignment = "Random", # distribution = "gaussian", # ntrees = 175, # seed = 123) #solomini.model <- h2o.gbm(x = 1:15, # y = 16, # training_frame = mini, # nfolds = 5, # fold_assignment = "Random", # distribution = "gaussian", # ntrees = 175, # seed = 123) # ### Duo #duo <- rbind(Bat[MT == "duo-fpp",], Bat[MT == "duo",]) # #md <- duo$matchDuration #set.seed(123) ; kmd <- kmeans(md, centers = 2, iter.max = 7000) ; duo$matchDuration <- NULL #duo$Bat_or_mini <- kmd$cluster # Bat : 2 , mini : 1 # ### I change data. If groupId same data, change data with feature's mean. ### Because winPlacePerc is same each team. Solo data isn't unnecessary this work. # #a <- duo$groupId #b <- factor(a, levels <- levels(factor(a)), order = T) #c <- as.numeric(b) ; duo$gi <- c ; duo$groupId = NULL # #f1 <- aggregate(duo$assists, by = list(duo$gi), mean) #f2 <- aggregate(duo$boosts, by = list(duo$gi), mean) #f3 <- aggregate(duo$damageDealt, by = list(duo$gi), mean) #f4 <- aggregate(duo$DBNOs, by = list(duo$gi), mean) #f5 <- aggregate(duo$heals, by = list(duo$gi), mean) #f6 <- aggregate(duo$killPlace, by = list(duo$gi), mean) #f7 <- aggregate(duo$killStreaks, by = list(duo$gi), mean) #f8 <- aggregate(duo$longestKill, by = list(duo$gi), mean) #f9 <- aggregate(duo$revives, by = list(duo$gi), mean) #f10 <- aggregate(duo$rideDistance, by = list(duo$gi), mean) #f11 <- aggregate(duo$swimDistance, by = list(duo$gi), mean) #f12 <- aggregate(duo$teamKills, by = list(duo$gi), mean) #f13 <- aggregate(duo$walkDistance, by = list(duo$gi), mean) #f14 <- aggregate(duo$weaponsAcquired, by = list(duo$gi), mean) #f15 <- aggregate(duo$rank, by = list(duo$gi), mean) # #duo <- duo[order(duo$gi), ] #a1 <- merge(duo, f1, by.x = "gi", by.y = "Group.1", all = T)$x #a2 <- merge(duo, f2, by.x = "gi", by.y = "Group.1", all = T)$x #a3 <- merge(duo, f3, by.x = "gi", by.y = "Group.1", all = T)$x #a4 <- merge(duo, f4, by.x = "gi", by.y = "Group.1", all = T)$x #a5 <- merge(duo, f5, by.x = "gi", by.y = "Group.1", all = T)$x #a6 <- merge(duo, f6, by.x = "gi", by.y = "Group.1", all = T)$x #a7 <- merge(duo, f7, by.x = "gi", by.y = "Group.1", all = T)$x #a8 <- merge(duo, f8, by.x = "gi", by.y = "Group.1", all = T)$x #a9 <- merge(duo, f9, by.x = "gi", by.y = "Group.1", all = T)$x #a10 <- merge(duo, f10, by.x = "gi", by.y = "Group.1", all = T)$x #a11 <- merge(duo, f11, by.x = "gi", by.y = "Group.1", all = T)$x #a12 <- merge(duo, f12, by.x = "gi", by.y = "Group.1", all = T)$x #a13 <- merge(duo, f13, by.x = "gi", by.y = "Group.1", all = T)$x #a14 <- merge(duo, f14, by.x = "gi", by.y = "Group.1", all = T)$x #a15 <- merge(duo, f15, by.x = "gi", by.y = "Group.1", all = T)$x # #duo$assists <- a1 ; duo$boosts <- a2 ; duo$damageDealt <- a3 ; duo$DBNOs <- a4 ; duo$heals <- a5 #duo$killPlace <- a6 ; duo$killStreaks <- a7 ; duo$longestKill <- a8 ; duo$revives <- a9 #duo$rideDistance <- a10 ; duo$swimDistance <- a11 ; duo$teamKills <- a12 ; duo$walkDistance <- a13 #duo$weaponsAcquired <- a14 ; duo$rank <- a15 #duo <- duo[order(as.integer(rownames(duo))), ] ; duo <- duo[!duplicated(duo$gi), ] #duo$gi <- NULL ; bom <- duo$Bat_or_mini ; duo$Bat_or_mini = NULL ; #bat <- duo[bom == 2, ] ; mini <- duo[bom == 1, ] # # #duo <- as.h2o(duo) ; bat <- as.h2o(bat) ; mini <- as.h2o(mini) #duobat.model <- h2o.gbm(x = 1:15, # y = 16, # training_frame = bat, # nfolds = 5, # fold_assignment = "Random", # ntrees = 175, # seed = 123) #duomini.model <- h2o.gbm(x = 1:15, # y = 16, # training_frame = mini, # fold_assignment ="Random", # ntrees = 175, # nfolds = 5, # seed = 123) # ### Squad, Flare and Crash Mode regressions are simliar duo data regrssion. ### So I'll pass those analysis. # ### Result #test <- read.csv("test_V2.csv") #solo <- rbind(test[MT == "solo", ], test[MT == "solo-fpp", ], test[MT == "normal-solo", ], test[MT == "normal-solo-fpp", ]) #duo <- rbind(test[MT == "duo", ], test[MT == "duo-fpp", ], test[MT == "normal-duo", ], test[MT == "normal-duo-fpp", ]) #squad <- rbind(test[MT == "squad", ], test[MT == "squad-fpp", ], test[MT == "normal-squad", ], test[MT == "normal-squad-fpp", ]) #flare <- rbind(test[MT == "flarefpp", ], test[MT == "flaretpp", ]) #crash <- rbind(test[MT == "crashfpp", ], test[MT == "crashtpp", ]) # #solobat <- as.h2o(solobat) ; solomini <- as.h2o(solomini) ; duobat <- as.h2o(duobat) ; duomini <- as.h2o(duomini) #squadbat <- as.h2o(squadbat) ; squadmini <- as.h2o(squadmini) ; flare <- as.h2o(flare) ; crash <- as.h2o(crash) # #solobatpredict <- h2o.predict(solobat.model, solobat) #solominipredict <- h2o.predict(solomini.model, solomini) #duobatpredict <- h2o.predict(duobat.model, duobat) #duominipredict <- h2o.predict(duomini.model, duomini) #squadbatpredict <- h2o.predict(squadbat.model, squadbat) #squadminipredict <- h2o.predict(squadmini.model, squadmini) #flarepredict <- h2o.predict(flare.model, flare) #crashpredict <- h2o.predict(crash.model, crash) # #solobat['per'] <- solobatpredict ; solomini['per'] <- solominipredict #duobat['per'] <- duobatpredict ; duomini['per'] <- duominipredict #squadbat['per'] <- squadbatpredict ; squadmini['per'] <- squadminipredict #flare['per'] <- flarepredict ; crash['per'] <- crashpredict # #solobat <- as.data.frame(solobat) #solomini <- as.data.frame(solomini) #duobat <- as.data.frame(duobat) #duomini <- as.data.frame(duomini) #squadbat <- as.data.frame(squadbat) #squadmini <- as.data.frame(squadmini) #flare <- as.data.frame(flare) #crash <- as.data.frame(crash) # #final_test <- rbind(solobat, solomini, duobat, duomini, squadbat, squadmini, flare, crash) #final_test <- final_test[order(final_test$nrow), ] #per <- final_test$per #per <- ifelse(per > 1, 1, per) #per <- ifelse(per < 0, 0, per) #final_test$per <- per # #sample_submission_V2 <- read.csv("sample_submission_V2.csv") #sample_submission_V2$winPlacePerc <- per #write.csv(sample_submission_V2, # file = "sample_submission_V2.csv", # row.names = T) # sub <- read.csv("../input/my-submission/sample_submission_V2.csv") sub$X = NULL write.csv(sub, file = "sample_submission.csv", row.names = F) # This work is first analysis in Kaggle. So I don't know how to submit... # I was able to submit it..