#Note: This is a work in progress. Any feedback or suggestions are greatly appreciated. #The flow of this kernel is different than typical kernels. Model training is still done first, followed by predicting test data, however objects are imported/created only when necessary and removed thereafter to minimize ram usage and allow for a larger training dataset. #Load necessary libraries. library(tidyverse) library(lightgbm) #Import training dataset. train_full <- read.csv("../input/train_V2.csv", header = T, stringsAsFactors = F) set.seed <- 42 #The answer to life, the Universe and everything. train <- train_full[sample(nrow(train_full), 3825000), ] #Sampling to test run-time. Choose a sample size, or nrow(train_full) for all training data. Max size currently 3.5M rows with only max and mean aggregations (incl. rank). Use 1M for exploring. rm(train_full) #TODO: EDA. #Feature engineering (some feature engineering is omitted (commented out) to preserve memory): #Players per match. playersPerMatch <- train %>% group_by(matchId) %>% summarize(playersPerMatch = n()) train <- merge(train, playersPerMatch, by = 'matchId', all.x = T) rm(playersPerMatch) #Party stats (party size; TBD). custom_group_stats <- train %>% group_by(matchId, groupId) %>% summarize(partySize = n()) train <- merge(train, custom_group_stats, by = c('matchId', 'groupId'), all.x = T) rm(custom_group_stats) #TODO: Add other engineered features here. #TODO: Flag cheaters and zombies. Remove? How to treat outliers. #Combination metrics (based on PUBG top kernels): #Sum of heals and boosts. #train <- mutate(train, healsAndBoosts = heals + boosts) #Could add weaponsAcquired to make this total items acquired (totalItemsAcquired). #Total distance travelled. #train <- mutate(train, totalDistance = walkDistance + rideDistance + swimDistance) #Teamwork metric. #train <- mutate(train, teamwork = assists + revives) #Sum of kills and assists. #train <- mutate(train, killsAndAssists = kills + assists) #Combat metric. #train <- mutate(train, combatPerf = damageDealt/100 + kills - teamKills + DBNOs + assists) #May be useful. #Group mean and max values per match. Common approach used in PUBG top kernels. Removed ', playersPerMatch, partySize' train_mean_max <- train %>% select(-c(winPlacePerc, playersPerMatch, partySize)) %>% group_by(matchId, groupId) %>% summarize_if(is.numeric, funs(mean, max)) #Only included max and mean aggregations under assumption weakest player's performance isn't important (i.e. leaders carry teams). #Create match rankings for each group. Ditto. train_mean_max_rank <- train_mean_max %>% group_by(matchId) %>% mutate(assists_mean_rank = rank(assists_mean, ties.method = 'min')/numGroups_max*100, #Mean boosts_mean_rank = rank(boosts_mean, ties.method = 'min')/numGroups_max*100, damageDealt_mean_rank = rank(damageDealt_mean, ties.method = 'min')/numGroups_max*100, DBNOs_mean_rank = rank(DBNOs_mean, ties.method = 'min')/numGroups_max*100, headshotKills_mean_rank = rank(headshotKills_mean, ties.method = 'min')/numGroups_max*100, heals_mean_rank = rank(heals_mean, ties.method = 'min')/numGroups_max*100, killPlace_mean_rank = rank(killPlace_mean, ties.method = 'min')/numGroups_max*100, #May not be necessary killPoints_mean_rank = rank(killPoints_mean, ties.method = 'min')/numGroups_max*100, #Ditto kills_mean_rank = rank(kills_mean, ties.method = 'min')/numGroups_max*100, killStreaks_mean_rank = rank(killStreaks_mean, ties.method = 'min')/numGroups_max*100, longestKill_mean_rank = rank(longestKill_mean, ties.method = 'min')/numGroups_max*100, matchDuration_mean_rank = rank(matchDuration_mean, ties.method = 'min')/numGroups_max*100, #Ditto maxPlace_mean_rank = rank(maxPlace_mean, ties.method = 'min')/numGroups_max*100, #Ditto numGroups_mean_rank = rank(numGroups_mean, ties.method = 'min')/numGroups_max*100, #Ditto rankPoints_mean_rank = rank(rankPoints_mean, ties.method = 'min')/numGroups_max*100, #Ditto revives_mean_rank = rank(revives_mean, ties.method = 'min')/numGroups_max*100, rideDistance_mean_rank = rank(rideDistance_mean, ties.method = 'min')/numGroups_max*100, roadKills_mean_rank = rank(roadKills_mean, ties.method = 'min')/numGroups_max*100, swimDistance_mean_rank = rank(swimDistance_mean, ties.method = 'min')/numGroups_max*100, teamKills_mean_rank = rank(teamKills_mean, ties.method = 'min')/numGroups_max*100, vehicleDestroys_mean_rank = rank(vehicleDestroys_mean, ties.method = 'min')/numGroups_max*100, walkDistance_mean_rank = rank(walkDistance_mean, ties.method = 'min')/numGroups_max*100, weaponsAcquired_mean_rank = rank(weaponsAcquired_mean, ties.method = 'min')/numGroups_max*100, winPoints_mean_rank = rank(winPoints_mean, ties.method = 'min')/numGroups_max*100, #Ditto #healsAndBoosts_mean_rank = rank(healsAndBoosts_mean, ties.method = 'min')/numGroups_max*100, #Custom #teamwork_mean_rank = rank(teamwork_mean, ties.method = 'min')/numGroups_max*100, #Custom #killsAndAssists_mean_rank = rank(killsAndAssists_mean, ties.method = 'min')/numGroups_max*100, #Custom #totalDistance_mean_rank = rank(totalDistance_mean, ties.method = 'min')/numGroups_max*100, #Custom #combatPerf_mean_rank = rank(combatPerf_mean, ties.method = 'min')/numGroups_max*100, #Custom assists_max_rank = rank(assists_max, ties.method = 'min')/numGroups_max*100, #Max boosts_max_rank = rank(boosts_max, ties.method = 'min')/numGroups_max*100, damageDealt_max_rank = rank(damageDealt_max, ties.method = 'min')/numGroups_max*100, DBNOs_max_rank = rank(DBNOs_max, ties.method = 'min')/numGroups_max*100, headshotKills_max_rank = rank(headshotKills_max, ties.method = 'min')/numGroups_max*100, heals_max_rank = rank(heals_max, ties.method = 'min')/numGroups_max*100, killPlace_max_rank = rank(killPlace_max, ties.method = 'min')/numGroups_max*100, #May not be necessary killPoints_max_rank = rank(killPoints_max, ties.method = 'min')/numGroups_max*100, #Ditto kills_max_rank = rank(kills_max, ties.method = 'min')/numGroups_max*100, killStreaks_max_rank = rank(killStreaks_max, ties.method = 'min')/numGroups_max*100, longestKill_max_rank = rank(longestKill_max, ties.method = 'min')/numGroups_max*100, matchDuration_max_rank = rank(matchDuration_max, ties.method = 'min')/numGroups_max*100, #Ditto maxPlace_max_rank = rank(maxPlace_max, ties.method = 'min')/numGroups_max*100, #Ditto numGroups_max_rank = rank(numGroups_max, ties.method = 'min')/numGroups_max*100, #Ditto rankPoints_max_rank = rank(rankPoints_max, ties.method = 'min')/numGroups_max*100, #Ditto revives_max_rank = rank(revives_max, ties.method = 'min')/numGroups_max*100, rideDistance_max_rank = rank(rideDistance_max, ties.method = 'min')/numGroups_max*100, roadKills_max_rank = rank(roadKills_max, ties.method = 'min')/numGroups_max*100, swimDistance_max_rank = rank(swimDistance_max, ties.method = 'min')/numGroups_max*100, teamkills_max_rank = rank(teamKills_max, ties.method = 'min')/numGroups_max*100, vehicleDestroys_max_rank = rank(vehicleDestroys_max, ties.method = 'min')/numGroups_max*100, walkDistance_max_rank = rank(walkDistance_max, ties.method = 'min')/numGroups_max*100, weaponsAcquired_max_rank = rank(weaponsAcquired_max, ties.method = 'min')/numGroups_max*100, winPoints_max_rank = rank(winPoints_max, ties.method = 'min')/numGroups_max*100, #Ditto #healsAndBoosts_max_rank = rank(healsAndBoosts_max, ties.method = 'min')/numGroups_max*100, #Custom #teamwork_max_rank = rank(teamwork_max, ties.method = 'min')/numGroups_max*100, #Custom #killsAndAssists_max_rank = rank(killsAndAssists_max, ties.method = 'min')/numGroups_max*100, #Custom #totalDistance_max_rank = rank(totalDistance_max, ties.method = 'min')/numGroups_max*100 #Custom #combatPerf_max_rank = rank(combatPerf_max, ties.method = 'min')/numGroups_max*100 #Custom ) rm(train_mean_max) gc() #Merging mean variables with original training dataset. train_prep <- merge(train, train_mean_max_rank, by = c('matchId', 'groupId'), all.x = T) rm(train_mean_max_rank) rm(train) #One-hot encoding of matchType. #TODO: Group all match types by solo, duo, squad (both custom and competitive). Could add flag for competitive or custom. matchType_train <- model.matrix(~matchType-1, train_prep) #Excluding Ids and matchType. train_prep <- train_prep %>% select(-c('Id', 'groupId', 'matchId', 'matchType')) #Merging in matchType dummy variables. train_prep <- cbind(train_prep, matchType_train) rm(matchType_train) #Preparing data for model training. output <- train_prep %>% na.omit() %>% select(winPlacePerc) train_prep <- train_prep %>% na.omit() %>% select(-winPlacePerc) %>% as.matrix() #Training LightGBM model with ad-hoc selection of parameters. #TODO: Incorporate grid search to find optimal parameters. #TODO: Incorporate stacked ensembling. #TODO: Incorporate model evaluation. #TODO: Explore removing unimportant features based on variable importance. #TODO: Explore PCA to reduce dimensionality and add more data. #train_prep.pca <- prcomp(train_prep[,apply(train_prep, 2, var, na.rm=TRUE) != 0], scale. = TRUE) #summary(train_prep.pca) gc() model <- lightgbm(data = train_prep, label = output$winPlacePerc, nrounds = 500, max_depth = 6, min_data_in_leaf = 5, #Or 3. early_stopping_round = 3, learning_rate = 0.1, objective = 'regression', metric='mae' ) rm(train_prep) rm(output) gc() #Repeat feature engineering for test dataset. test <- read.csv("../input/test_V2.csv", header = T, stringsAsFactors = F) #Players per match. playersPerMatch <- test %>% group_by(matchId) %>% summarize(playersPerMatch = n()) test <- merge(test, playersPerMatch, by = 'matchId', all.x = T) rm(playersPerMatch) #Party stats (party size; TBD). custom_group_stats <- test %>% group_by(matchId, groupId) %>% summarize(partySize = n()) test <- merge(test, custom_group_stats, by = c('matchId', 'groupId'), all.x = T) rm(custom_group_stats) #Sum of heals and boosts. #test <- mutate(test, healsAndBoosts = heals + boosts) #Total distance travelled. #test <- mutate(test, totalDistance = walkDistance + rideDistance + swimDistance) #Teamwork metric. #test <- mutate(test, teamwork = assists + revives) #Sum of kills and assists. #test <- mutate(test, killsAndAssists = kills + assists) #Combat metric. #test <- mutate(test, combatPerf = damageDealt/100 + kills - teamKills + DBNOs + assists) #Group mean, min, max values per match. Removed '%>% select(-c(playersPerMatch, partySize))' test_mean_max <- test %>% select(-c(playersPerMatch, partySize)) %>% group_by(matchId, groupId) %>% summarize_if(is.numeric, funs(mean, max)) #Create match rankings for each group. test_mean_max_rank <- test_mean_max %>% group_by(matchId) %>% mutate(assists_mean_rank = rank(assists_mean, ties.method = 'min')/numGroups_max*100, #Mean boosts_mean_rank = rank(boosts_mean, ties.method = 'min')/numGroups_max*100, damageDealt_mean_rank = rank(damageDealt_mean, ties.method = 'min')/numGroups_max*100, DBNOs_mean_rank = rank(DBNOs_mean, ties.method = 'min')/numGroups_max*100, headshotKills_mean_rank = rank(headshotKills_mean, ties.method = 'min')/numGroups_max*100, heals_mean_rank = rank(heals_mean, ties.method = 'min')/numGroups_max*100, killPlace_mean_rank = rank(killPlace_mean, ties.method = 'min')/numGroups_max*100, #May not be necessary killPoints_mean_rank = rank(killPoints_mean, ties.method = 'min')/numGroups_max*100, #Ditto kills_mean_rank = rank(kills_mean, ties.method = 'min')/numGroups_max*100, killStreaks_mean_rank = rank(killStreaks_mean, ties.method = 'min')/numGroups_max*100, longestKill_mean_rank = rank(longestKill_mean, ties.method = 'min')/numGroups_max*100, matchDuration_mean_rank = rank(matchDuration_mean, ties.method = 'min')/numGroups_max*100, #Ditto maxPlace_mean_rank = rank(maxPlace_mean, ties.method = 'min')/numGroups_max*100, #Ditto numGroups_mean_rank = rank(numGroups_mean, ties.method = 'min')/numGroups_max*100, #Ditto rankPoints_mean_rank = rank(rankPoints_mean, ties.method = 'min')/numGroups_max*100, #Ditto revives_mean_rank = rank(revives_mean, ties.method = 'min')/numGroups_max*100, rideDistance_mean_rank = rank(rideDistance_mean, ties.method = 'min')/numGroups_max*100, roadKills_mean_rank = rank(roadKills_mean, ties.method = 'min')/numGroups_max*100, swimDistance_mean_rank = rank(swimDistance_mean, ties.method = 'min')/numGroups_max*100, teamKills_mean_rank = rank(teamKills_mean, ties.method = 'min')/numGroups_max*100, vehicleDestroys_mean_rank = rank(vehicleDestroys_mean, ties.method = 'min')/numGroups_max*100, walkDistance_mean_rank = rank(walkDistance_mean, ties.method = 'min')/numGroups_max*100, weaponsAcquired_mean_rank = rank(weaponsAcquired_mean, ties.method = 'min')/numGroups_max*100, winPoints_mean_rank = rank(winPoints_mean, ties.method = 'min')/numGroups_max*100, #Ditto #healsAndBoosts_mean_rank = rank(healsAndBoosts_mean, ties.method = 'min')/numGroups_max*100, #Custom #teamwork_mean_rank = rank(teamwork_mean, ties.method = 'min')/numGroups_max*100, #Custom #killsAndAssists_mean_rank = rank(killsAndAssists_mean, ties.method = 'min')/numGroups_max*100, #Custom #totalDistance_mean_rank = rank(totalDistance_mean, ties.method = 'min')/numGroups_max*100, #Custom #combatPerf_mean_rank = rank(combatPerf_mean, ties.method = 'min')/numGroups_max*100, #Custom assists_max_rank = rank(assists_max, ties.method = 'min')/numGroups_max*100, #Max boosts_max_rank = rank(boosts_max, ties.method = 'min')/numGroups_max*100, damageDealt_max_rank = rank(damageDealt_max, ties.method = 'min')/numGroups_max*100, DBNOs_max_rank = rank(DBNOs_max, ties.method = 'min')/numGroups_max*100, headshotKills_max_rank = rank(headshotKills_max, ties.method = 'min')/numGroups_max*100, heals_max_rank = rank(heals_max, ties.method = 'min')/numGroups_max*100, killPlace_max_rank = rank(killPlace_max, ties.method = 'min')/numGroups_max*100, #May not be necessary killPoints_max_rank = rank(killPoints_max, ties.method = 'min')/numGroups_max*100, #Ditto kills_max_rank = rank(kills_max, ties.method = 'min')/numGroups_max*100, killStreaks_max_rank = rank(killStreaks_max, ties.method = 'min')/numGroups_max*100, longestKill_max_rank = rank(longestKill_max, ties.method = 'min')/numGroups_max*100, matchDuration_max_rank = rank(matchDuration_max, ties.method = 'min')/numGroups_max*100, #Ditto maxPlace_max_rank = rank(maxPlace_max, ties.method = 'min')/numGroups_max*100, #Ditto numGroups_max_rank = rank(numGroups_max, ties.method = 'min')/numGroups_max*100, #Ditto rankPoints_max_rank = rank(rankPoints_max, ties.method = 'min')/numGroups_max*100, #Ditto revives_max_rank = rank(revives_max, ties.method = 'min')/numGroups_max*100, rideDistance_max_rank = rank(rideDistance_max, ties.method = 'min')/numGroups_max*100, roadKills_max_rank = rank(roadKills_max, ties.method = 'min')/numGroups_max*100, swimDistance_max_rank = rank(swimDistance_max, ties.method = 'min')/numGroups_max*100, teamkills_max_rank = rank(teamKills_max, ties.method = 'min')/numGroups_max*100, vehicleDestroys_max_rank = rank(vehicleDestroys_max, ties.method = 'min')/numGroups_max*100, walkDistance_max_rank = rank(walkDistance_max, ties.method = 'min')/numGroups_max*100, weaponsAcquired_max_rank = rank(weaponsAcquired_max, ties.method = 'min')/numGroups_max*100, winPoints_max_rank = rank(winPoints_max, ties.method = 'min')/numGroups_max*100, #Ditto #healsAndBoosts_max_rank = rank(healsAndBoosts_max, ties.method = 'min')/numGroups_max*100, #Custom #teamwork_max_rank = rank(teamwork_max, ties.method = 'min')/numGroups_max*100, #Custom #killsAndAssists_max_rank = rank(killsAndAssists_max, ties.method = 'min')/numGroups_max*100, #Custom #totalDistance_max_rank = rank(totalDistance_max, ties.method = 'min')/numGroups_max*100 #Custom #combatPerf_max_rank = rank(combatPerf_max, ties.method = 'min')/numGroups_max*100 #Custom ) rm(test_mean_max) #Merging mean variables with original test dataset. test_prep <- merge(test, test_mean_max_rank, by = c('matchId', 'groupId'), all.x = T) rm(test_mean_max_rank) rm(test) #One-hot encoding of matchType. matchType_test <- model.matrix(~matchType-1, test_prep) #Excluding Ids and matchType. test.Id <- test_prep$Id test_prep <- test_prep %>% select(-c('Id', 'groupId', 'matchId', 'matchType')) #Merging in matchType dummy variables. test_prep <- cbind(test_prep, matchType_test) rm(matchType_test) #Generate predictions on test dataset. test_prep$winPlacePerc.predict <- predict(model, as.matrix(test_prep)) submission <- data.frame(Id = test.Id, winPlacePerc = test_prep$winPlacePerc.predict, stringsAsFactors = F) submission$winPlacePerc <- ifelse(submission$winPlacePerc < 0, 0, ifelse(submission$winPlacePerc > 1, 1, submission$winPlacePerc)) write.csv(submission, "submission_lgbm_nrow3.75M_nrounds500.csv", row.names = F)