[R 프로그래밍] 머신 러닝

# 중간 데이터 읽기용 함수 작성

library(plyr)

library(foreach)

readActionHourly <- function(app.name, date.from, date.to) {

date.from <- as.Date(date.from)

date.to <- as.Date(date.to)

ldply(foreach(day = seq.Date(date.from, date.to, by = "day"),

combine = rbind) %do%

{

f <- sprintf("sample-data/section10/action_hourly/%s/%s/action_hourly.tsv", app.name, day)

read.csv(f, header = T, stringsAsFactors = F, sep = "\t")

})

}

# 작성한 함수로 데이터 읽어들이기

action.hourly <- readActionHourly("game-01", "2013-08-01", "2013-08-08")

head(action.hourly)

# 시간대가 열에 오도록 정렬

library(reshape2)

dates <- unique(action.hourly$log_date)

train.list <- lapply(1:(length(dates) - 1), function(i) {

day <- dates[i]

x <- action.hourly[action.hourly$log_date == day, ]

df <- dcast(x, user_id ~ log_hour, value.var = "count",

function(x) {ifelse(sum(x) >= 7, 1, 0)}

)

names(df) <- c("user_id", paste0("p", i, "_", 0:23))

df})

head(train.list[[1]])

# 설명용 데이터 작성

train.data <- train.list[[1]]

for (i in 2:length(train.list)) {

df <- train.list[[i]]

train.data <- merge(train.data, df, by = "user_id", all.x = T)

train.data[is.na(train.data)] <- 0

}

names(train.data)

# 정답 데이터 작성

ans0 <- action.hourly[action.hourly$log_date == dates[length(dates)], ]

# 시간대가 열에 오도록 정렬

ans <- dcast(ans0, user_id ~ log_hour, value.var = "count", function(x) {

ifelse(sum(x) >= 7, 1, 0)

})

names(ans) <- c("user_id", paste0("a_", 0:23))

head(ans)

# 설명용 데이터와 결합하기

train.data <- merge(train.data, ans, by = "user_id", all.x = T)

train.data[is.na(train.data)] <- 0

names(train.data)

# 시간대별로 모델 구축(랜덤포레스트)

library(randomForest)

library(caret)

# 학습 조절

fit.control <- trainControl(method = "LGOCV", p = 0.75, number = 30)

# 랜덤포레스트에 의한 학습실시

rf.fit.list <- lapply(0:23, function(h) {

df <- train.data[, c(paste0("p", 1:7, "_", h), paste0("a_", h))]

df1 <- df[df[, ncol(df)] == 1, ]

df0.orig <- df[df[, ncol(df)] == 0, ]

df0 <- df0.orig[sample(1:nrow(df0.orig), nrow(df1)), ]

df <- rbind(df1, df0)

fit <- train(x=df[, -ncol(df)], y=df[, ncol(df)],

method = "rf", preProcess = c("center", "scale"),

trControl = fit.control)

fit

})

# 작성한 모델의 예측치와 실제값의 비교 결과를 표시

for (i in 0:23) {

result.rf <- predict(rf.fit.list[[(i + 1)]]$finalModel, train.data)

result.rf.pred <- ifelse(result.rf >= 0.5, 1, 0)

print(confusionMatrix(result.rf.pred, train.data[, paste0("a_",

i)]))

}