#Data visualization
cars <- read.table("Cars.csv", header = TRUE, sep=";")
plot(cars$cena~cars$rok_vyroby)
plot(cars$objem~cars$vykon)
boxplot(cars$cena~cars$vyrobce, las = 2)
table(cars$vyrobce)




vyrobcee <- c(mdt2$vyrobce)

plot(mdt$rating~mdt$vyrobce, las = 2)
plot(mdt$rating~mdt$model, las = 2)
plot(mdt$rating~mdt$rok_vyroby)
plot(mdt$rating~mdt$najete_km)
plot(mdt$rating~mdt$objem)
plot(mdt$rating~mdt$vykon)
plot(mdt$rating~mdt$spotreba)
plot(mdt$rating~mdt$cena)

abline(lm(mdt$rating~mdt$vykon))

#
dt1 <- c(0,1,2,3,4,5,6,7,8,9)
dt2 <- c(0,1,2,3,3.4,3.9,5,4.5,3,2)
df <- data.frame(dt1, dt2)

#data discretization
library(arules)
discretize(dt1)
discretize(dt1, method="frequency", categories=4)

#various regression methods
library(earth)

mod <- earth(dt2~dt1, df)
mod2 <- lm(dt2~dt1, df)

summary(mod)
summary(mod2)

plot(c(0,9), c(0,5), type = "n")
plotmo(mod)
points(dt1, dt2, type = "p")
lines(dt1,dt2 ,col="green")

plot(c(0,9), c(0,5), type = "n")
abline(mod2)
points(dt1, dt2, type = "p")
lines(dt1,dt2 ,col="green")

peak <- which.max(df$dt2)
lm1 <- lm(dt2~dt1, df[0:peak,])
lm2 <- lm(dt2~dt1, df[peak:nrow(df),])
plot(c(0,9), c(0,5), type = "n")
abline(lm1)
abline(lm2)
points(dt1, dt2, type = "p")


peak <- which.max(df$dt2)
plot(c(0,9), c(0,5), type = "n")
points(c(df$dt1[1], df$dt1[peak], df$dt1[nrow(df)]) , c(df$dt2[1], df$dt2[peak], df$dt2[nrow(df)]), type = "l")
points(dt1, dt2, type = "p")

# error on the train set

mae(predict(mod, newdata = df), dt2)
mae(predict(mod2, newdata = df), dt2)

mod2 <- lm(dt2~dt1, df)

predict(mod2, data.frame(dt1 = c(1.5, 5, 8)))

model <- lm(dt2 ~ dt1, data=df)
new.df <- data.frame(dt1=c(7, 8, 4))
wtk <- predict(model, new.df)
dfd <- data.frame(new.df, wtf)

myVar = 6;
if(myVar<peak){
predict(lm1, newdata = data.frame(dt1=c(myVar)))
} else {
predict(lm2, newdata = data.frame(dt1=c(myVar)))
}

dt3 <- c(
  predict(lm1, newdata = df[0:peak,]),
  predict(lm2, newdata = df[(peak+1):nrow(df),])
  )
mae(dt3, dt2)

rmse(dt3, dt2)

dtt1 <- c(7,4,6,9)
dtt2 <- c(1,3,5,7)
dft <- data.frame(dtt1, dtt2)

m <- cbind(predict(mod2, newdata = dft), dtt2)
cor(m, method="kendall", use="pairwise")




# error on the test set (your task) is more important due to overfitting

###############################   HOMEWORK   #############################################
library(Metrics)


FINAL_RESULTS <- data.frame(uid = integer(0), oid= integer(0), estimated_rating = integer(0))


for (i in 1:100) {
    train <- read.table("Train.csv", header = TRUE, sep=";")
    train <- train[which(train$uid==i),]

    test <- read.table("Test.csv", header = TRUE, sep=";")
    test <- test[which(test$uid==i),]

    mdt <- merge(cars, train, by.x="id", by.y="oid")
    mdt2 <- merge(cars, test, by.x="id", by.y="oid")

######## 1) REGRESSIONS FOR NUMERICAL :

#1 Vykon

vdt1 <- c(mdt$vykon)
vdt2 <- c(mdt$rating)
vdf <- data.frame(vdt1, vdt2)

vdf <- vdf[order(vdt1),]
vpeak <- which.max(vdf$vdt2)
vlm1 <- lm(vdt2~vdt1, vdf[which(vdf$vdt1 <= vdf$vdt1[vpeak]),])
vlm2 <- lm(vdt2~vdt1, vdf[which(vdf$vdt1 >= vdf$vdt1[vpeak]),])
plot(mdt$rating~mdt$vykon)
if(is.na(vlm2$coefficients[2])){ #HERE!!!
  vlm2=vlm1
}
if(is.na(vlm1$coefficients[2])){
  vlm1=vlm2
}
abline(vlm1)
abline(vlm2)
points(vdt1, vdt2, type = "p")

# vpeak <- which.max(vdf$vdt2)
# plot(mdt$rating~mdt$vykon)
# points(c(vdf$vdt1[1], vdf$vdt1[vpeak], vdf$vdt1[nrow(vdf)]) , c(vdf$vdt2[1], vdf$vdt2[vpeak], vdf$vdt2[nrow(vdf)]), type = "l")
# points(vdt1, vdt2, type = "p")

#2 Rok wyroby

rwdt1 <- c(mdt$rok_vyroby)
rwdt2 <- c(mdt$rating)
rwdf <- data.frame(rwdt1, rwdt2)

rwdf <- rwdf[order(rwdt1),]

rwpeak <- which.max(rwdf$rwdt2)
rwlm1 <- lm(rwdt2~rwdt1, rwdf[which(rwdf$rwdt1 <= rwdf$rwdt1[rwpeak]),])
rwlm2 <- lm(rwdt2~rwdt1, rwdf[which(rwdf$rwdt1 >= rwdf$rwdt1[rwpeak]),])
plot(mdt$rating~mdt$rok_vyroby)
if(is.na(rwlm2$coefficients[2])){
  rwlm2=rwlm1
}
if(is.na(rwlm1$coefficients[2])){
  rwlm1=rwlm2
}
abline(rwlm1)
abline(rwlm2)
points(rwdt1, rwdt2, type = "p")

# rwpeak <- which.max(rwdf$rwdt2)
# plot(mdt$rating~mdt$rok_vyroby)
# points(c(rwdf$rwdt1[1], rwdf$rwdt1[rwpeak], rwdf$rwdt1[nrow(rwdf)]) , c(rwdf$rwdt2[1], rwdf$rwdt2[rwpeak], rwdf$rwdt2[nrow(rwdf)]), type = "l")
# points(rwdt1, rwdt2, type = "p")

#3 Najete km

ndt1 <- c(mdt$najete_km)
ndt2 <- c(mdt$rating)
ndf <- data.frame(ndt1, ndt2)

ndf <- ndf[order(ndt1),]

npeak <- which.max(ndf$ndt2)
nlm1 <- lm(ndt2~ndt1, ndf[which(ndf$ndt1 <= ndf$ndt1[npeak]),])
nlm2 <- lm(ndt2~ndt1, ndf[which(ndf$ndt1 >= ndf$ndt1[npeak]),])
plot(mdt$rating~mdt$najete_km)
if(is.na(nlm2$coefficients[2])){
  nlm2=nlm1
}
if(is.na(nlm1$coefficients[2])){
  nlm1=nlm2
}
abline(nlm1)
abline(nlm2)
points(ndt1, ndt2, type = "p")

# npeak <- which.max(ndf$ndt2)
# plot(mdt$rating~mdt$najete_km)
# points(c(ndf$ndt1[1], ndf$ndt1[npeak], ndf$ndt1[nrow(ndf)]) , c(ndf$ndt2[1], ndf$ndt2[npeak], ndf$ndt2[nrow(ndf)]), type = "l")
# points(ndt1, ndt2, type = "p")

#4 Objem

odt1 <- c(mdt$objem)
odt2 <- c(mdt$rating)
odf <- data.frame(odt1, odt2)

odf <- odf[order(odt1),]

opeak <- which.max(odf$odt2)
olm1 <- lm(odt2~odt1, odf[which(odf$odt1 <= odf$odt1[opeak]),])
olm2 <- lm(odt2~odt1, odf[which(odf$odt1 >= odf$odt1[opeak]),])
plot(mdt$rating~mdt$objem)
if(is.na(olm2$coefficients[2])){
  olm2=olm1
}
if(is.na(olm1$coefficients[2])){
  olm1=olm2
}
abline(olm1)
abline(olm2)
points(odt1, odt2, type = "p")

# opeak <- which.max(odf$odt2)
# plot(mdt$rating~mdt$objem)
# points(c(odf$odt1[1], odf$odt1[opeak], odf$odt1[nrow(odf)]) , c(odf$odt2[1], odf$odt2[opeak], odf$odt2[nrow(odf)]), type = "l")
# points(odt1, odt2, type = "p")

#5 Spotreba

sdt1 <- c(mdt$spotreba)
sdt2 <- c(mdt$rating)
sdf <- data.frame(sdt1, sdt2)

sdf <- sdf[order(sdt1),]

speak <- which.max(sdf$sdt2)
slm1 <- lm(sdt2~sdt1, sdf[which(sdf$sdt1 <= sdf$sdt1[speak]),])
slm2 <- lm(sdt2~sdt1, sdf[which(sdf$sdt1 >= sdf$sdt1[speak]),])
plot(mdt$rating~mdt$spotreba)
if(is.na(slm2$coefficients[2])){
  slm2=slm1
}
if(is.na(slm1$coefficients[2])){
  slm1=slm2
}
abline(slm1)
abline(slm2)
points(sdt1, sdt2, type = "p")

# speak <- which.max(sdf$sdt2)
# plot(mdt$rating~mdt$spotreba)
# points(c(sdf$sdt1[1], sdf$sdt1[speak], sdf$sdt1[nrow(sdf)]) , c(sdf$sdt2[1], sdf$sdt2[speak], sdf$sdt2[nrow(sdf)]), type = "l")
# points(sdt1, sdt2, type = "p")

#6 Cena

cdt1 <- c(mdt$cena)
cdt2 <- c(mdt$rating)
cdf <- data.frame(cdt1, cdt2)

cdf <- cdf[order(cdt1),]

cpeak <- which.max(cdf$cdt2)
clm1 <- lm(cdt2~cdt1, cdf[which(cdf$cdt1 <= cdf$cdt1[cpeak]),])
clm2 <- lm(cdt2~cdt1, cdf[which(cdf$cdt1 >= cdf$cdt1[cpeak]),])
plot(mdt$rating~mdt$cena)
if(is.na(clm2$coefficients[2])){
  clm2=clm1
}
if(is.na(clm1$coefficients[2])){
  clm1=clm2
}
abline(clm1)
abline(clm2)
points(cdt1, cdt2, type = "p")

# cpeak <- which.max(cdf$cdt2)
# plot(mdt$rating~mdt$cena)
# points(c(cdf$cdt1[1], cdf$cdt1[cpeak], cdf$cdt1[nrow(cdf)]) , c(cdf$cdt2[1], cdf$cdt2[cpeak], cdf$cdt2[nrow(cdf)]), type = "l")
# points(cdt1, cdt2, type = "p")

########### 2) AVERAGES FOR NOMINAL:
vqdt2 <- c(mdt$rating)
vqdt3 <- c(mdt$model)
vqdt4 <- c(mdt$vyrobce)

vqdf2 <- data.frame(vqdt3, vqdt2)
vqdf3 <- data.frame(vqdt4, vqdt2)

#1 Model
agrmod <- aggregate(.~mdt$model, data=vqdf2, mean)

#2 Vyrobce
agrvyr <- aggregate(.~mdt$vyrobce, data=vqdf3, mean)

######### 3) ERRORS + WEIGHTS ON NUMERICAL

#Vykon peak:
vdt3 <- c(
  predict(vlm1, newdata = vdf[which(vdf$vdt1 < vdf$vdt1[vpeak]),]),
  predict(vlm2, newdata = vdf[which(vdf$vdt1 >= vdf$vdt1[vpeak]),])
)
mae(vdt3, vdt2)
vweight <- 1/rmse(vdt3, vdt2)

#Rok wyroby peak:
rwdt3 <- c(
  predict(rwlm1, newdata = rwdf[which(rwdf$rwdt1 < rwdf$rwdt1[rwpeak]),]),
  predict(rwlm2, newdata = rwdf[which(rwdf$rwdt1 >= rwdf$rwdt1[rwpeak]),])
)
mae(rwdt3, rwdt2)
rwweight <- 1/rmse(rwdt3, rwdt2)

#objem peak:
odt3 <- c(
  predict(olm1, newdata = odf[which(odf$odt1 < odf$odt1[opeak]),]),
  predict(olm2, newdata = odf[which(odf$odt1 >= odf$odt1[opeak]),])
)
mae(odt3, odt2)
oweight <- 1/rmse(odt3, odt2)


#Cena peak:
cdt3 <- c(
  predict(clm1, newdata = cdf[which(cdf$cdt1 < cdf$cdt1[cpeak]),]),
  predict(clm2, newdata = cdf[which(cdf$cdt1 >= cdf$cdt1[cpeak]),])
)
mae(cdt3, cdt2)
cweight <- 1/rmse(cdt3, cdt2)


#Spotreba peak
sdt3 <- c(
  predict(slm1, newdata = sdf[which(sdf$sdt1 < sdf$sdt1[speak]),]),
  predict(slm2, newdata = sdf[which(sdf$sdt1 >= sdf$sdt1[speak]),])
)
mae(sdt3, sdt2)
sweight <- 1/rmse(sdt3, sdt2)


#najete km
ndt3 <- c(
  predict(nlm1, newdata = ndf[which(ndf$ndt1 < ndf$ndt1[npeak]),]),
  predict(nlm2, newdata = ndf[which(ndf$ndt1 >= ndf$ndt1[npeak]),])
)
mae(ndt3, ndt2)
nweight <- 1/rmse(ndt3, ndt2)

##### 4) ERRORS + WEIGHTS ON NOMINAL

#weight of vyrobice:
barbar <- merge(mdt, agrvyr, by.x="vyrobce", by.y="mdt$vyrobce")
vyweight <- 1/rmse(barbar$rating,barbar$vqdt2)

#weight of model:

birbir <- merge(mdt, agrmod, by.x="model", by.y="mdt$model")
mweight <- 1/rmse(birbir$rating,birbir$vqdt2)



###### 5) PREDICTING NUMERICAL ATRIBUTES VALUES

tvykon <- c(mdt2$vykon)
tvdf <- data.frame(tvykon,c(1:49))
colnames(tvdf) <- colnames(vdf)
tvdt3 <- c(
  predict(vlm1, newdata = tvdf[which(tvdf$vdt1 < vdf$vdt1[vpeak]),]),
  predict(vlm2, newdata = tvdf[which(tvdf$vdt1 >= vdf$vdt1[vpeak]),])
)

tcena <- c(mdt2$cena)
tcdf <- data.frame(tcena,c(1:49))
colnames(tcdf) <- colnames(cdf)
tcdt3 <- c(
  predict(clm1, newdata = tcdf[which(tcdf$cdt1 < cdf$cdt1[cpeak]),]),
  predict(clm2, newdata = tcdf[which(tcdf$cdt1 >= cdf$cdt1[cpeak]),])
)

tobjem <- c(mdt2$objem)
todf <- data.frame(tobjem, c(1:49))
colnames(todf) <- colnames(odf)
todt3 <- c(
  predict(olm1, newdata = todf[which(todf$odt1 < odf$odt1[opeak]),]),
  predict(olm2, newdata = todf[which(todf$odt1 >= odf$odt1[opeak]),])
)

trokvyroby <- c(mdt2$rok_vyroby)
trwdf <- data.frame(trokvyroby, c(1:49))
colnames(trwdf) <- colnames(rwdf)
trwdt3 <- c(
  predict(rwlm1, newdata = trwdf[which(trwdf$rwdt1 < rwdf$rwdt1[rwpeak]),]),
  predict(rwlm2, newdata = trwdf[which(trwdf$rwdt1 >= rwdf$rwdt1[rwpeak]),])
)

tnajeteKm <- c(mdt2$najete_km)
tndf <- data.frame(tnajeteKm, c(1:49))
colnames(tndf) <- colnames(ndf)
tndt3 <- c(
  predict(nlm1, newdata = tndf[which(tndf$ndt1 < ndf$ndt1[npeak]),]),
  predict(nlm2, newdata = tndf[which(tndf$ndt1 >= ndf$ndt1[npeak]),])
)

tspotreba <- c(mdt2$spotreba)
tsdf <- data.frame(tspotreba, c(1:49))
colnames(tsdf) <- colnames(sdf)
tsdt3 <- c(
  predict(slm1, newdata = tsdf[which(tsdf$sdt1 < sdf$sdt1[speak]),]),
  predict(slm2, newdata = tsdf[which(tsdf$sdt1 >= sdf$sdt1[speak]),])
)


####### 6) PREDICTING NOMINAL ATRIBUTES VALUES

#(replacing NA's with averages)
tvyrobce <- mdt2$vyrobce
byrbyr <- merge(mdt2, agrvyr, by.x="vyrobce", by.y="mdt$vyrobce")
tvydf <- c(byrbyr$vqdt2)
meanTvydf  <- mean(tvydf)
byrbyr <- merge(mdt2, agrvyr, by.x="vyrobce", by.y="mdt$vyrobce", all.x = TRUE)
tvydf <- c(byrbyr$vqdt2)
tvydf[is.na(tvydf)] <- meanTvydf

tmodel <- mdt2$model
byrbyr2 <- merge(mdt2, agrmod, by.x="model", by.y="mdt$model")
tmdf <- c(byrbyr2$vqdt2)
meanTmf  <- mean(tmdf)
byrbyr2 <- merge(mdt2, agrmod, by.x="model", by.y="mdt$model", all.x = TRUE)
tmdf <- c(byrbyr2$vqdt2)
tmdf[is.na(tmdf)] <- meanTmf

####### 7) AGGREGATING WEIGHTED VALUES TO PREDICT RATINGS ON TEST SET

#without weights:
rsNoWeights <- (tvdt3 + tcdt3 + todt3+ tsdt3 + tndt3 + trwdt3 + tvydf +tmdf)/8

weightsum <- vweight+cweight+oweight+sweight+nweight+rwweight+vyweight+mweight

results <- (tvdt3 *vweight + tcdt3 *cweight + todt3*oweight + tsdt3*sweight + tndt3*nweight + trwdt3*rwweight + tvydf*vyweight +tmdf*mweight)/weightsum
RESULT_SET <- data.frame(i,mdt2$id,results)
FINAL_RESULTS <- rbind(FINAL_RESULTS,RESULT_SET)
}
colnames(FINAL_RESULTS)[1] <- "uid"
colnames(FINAL_RESULTS)[2] <- "oid"
colnames(FINAL_RESULTS)[3] <- "estimated_rating"

write.table(FINAL_RESULTS, file = "Results.csv", row.names = FALSE)