HK Pollution Data Exploration

data2016 <- read.csv("../data/KC2016_04.csv")
head(data2016)

#Data type conversion and pulling out date parts into new columns
data2016$date <- as.POSIXct(as.character(data2016$date ), format = "%d/%m/%Y %H:%M") 
data2016$m <- paste(format(data2016$date, format="%m"),"_",format(data2016$date, format="%B"), sep="")
data2016$d <- paste(format(data2016$date, format="%u"),"_",format(data2016$date, format="%A"), sep="") 
data2016$t <- format(data2016$date, format="%R")
data2016$FSP <- as.numeric(data2016$FSP)
head(data2016[,8:13])

2016 Pollution By Month

p <- ggplot(data2016, aes(m, FSP, fill=m)) + geom_boxplot() 
p

Prepare Training and Test Sets

# remove bad rows and unused cols
data2016 <- data2016[!is.na(data2016$date),]
# Create dummy vars
data2016 <- cbind(data2016, 
                  model.matrix(~ m -1, data2016), 
                  model.matrix(~ d -1, data2016),
                  model.matrix(~ t -1, data2016))
data2016mdl <- data2016[,c(10,14:56)]
names(data2016mdl) <- gsub(":00","",names(data2016mdl))
# Split test and train
sample <- sample.int(n = nrow(data2016mdl), size = floor(.90 * nrow(data2016mdl)), replace = F)
data2016.train <- data2016mdl[sample, ]
data2016.test  <- data2016mdl[-sample, ]
head(data2016.train)

Build Decision Tree Model and Predict

dt.mdl <- rpart(FSP ~ .,data=data2016.train,control=rpart.control(minsplit=1, minbucket=1, cp=0.0015))
plot(dt.mdl)
text(dt.mdl,cex=0.5)

data2016.test$predFSP<-predict(dt.mdl,data2016.test) 

Prediction Accuracy (Delta Actual FSP & Predicted FSP)

bins <- 20
cols <- c("green","orange", "red", "darkred")
colGradient <- colorRampPalette(cols)
cut.cols <- colGradient(bins)
cuts <- cut(data2016.test$FSP,bins)
names(cuts) <- sapply(cuts,function(t) cut.cols[which(as.character(t) == levels(cuts))])
p <- ggplot(data2016.test, aes(abs(FSP-predFSP),fill=cut(abs(FSP-predFSP),bins))) + 
  geom_histogram(binwidth=1,show.legend = FALSE) + 
  scale_color_manual(values=cut.cols,labels=levels(cuts)) +
  scale_fill_manual(values=cut.cols,labels=levels(cuts)) 
p