rossmann-store-sales

competitions/rossmann-store-sales/TODO_Rossmann

@@ -0,0 +1,35 @@
+
+** train / test (merge with store)
+- Sales - response 
+- Customers - cut 
+
+- Store 
+- DayOfWeek       
+- Date ~ dateNum , day , month , year     
+- Open 
+- Promo 
+- StateHoliday 
+- SchoolHoliday
+
+- StoreType 
+- Assortment 
+- CompetitionDistancePred := -1 if Date < CompetitionOpenSinceDate | CompetitionDistance is NA, otherwise CompetitionDistance (even if CompetitionOpenSinceDate is NA)
+- Promo2Pred := 0 if Promo2 is 0 | Promo2SinceDate < Date | Date not in Promo2Period , 1 otherwise
+- Promo2SinceMonths := # months that Promo2 started 
+
+* evaluate (add these predictors one by one)
+- SalesLastYearsDay := Sales of last years in the same shop
+- SalesLastYearWeek := Average Sales of last year in the same week 
+
+** feature selection 
+- Open = 1 
+- Sales >0 ?  
+
+** metric 
+- RMSPE : not takes into account Sales == 0 in days Open  (54 cases Shop / day , 41 stores)
+- RMSE  : takes into account Sales == 0 in days Open
+
+** resampling 
+- it seems that train period is before test period (~ 1,5 months)
+
+

competitions/rossmann-store-sales/bech.R

@@ -0,0 +1,139 @@
+library(readr)
+library(xgboost)
+library(fastfurious)
+
+### FAST-FURIOUS 
+ff.setBasePath(path = '/Users/gino/kaggle/fast-furious/gitHub/fast-furious/')
+ff.bindPath(type = 'data' , sub_path = 'dataset/rossmann-store-sales')
+ff.bindPath(type = 'code' , sub_path = 'competitions/rossmann-store-sales')
+ff.bindPath(type = 'elab' , sub_path = 'dataset/rossmann-store-sales/elab') 
+
+ff.bindPath(type = 'ensemble_1' , sub_path = 'dataset/rossmann-store-sales/ensembles/ensemble_1',createDir = T) ## out 
+ff.bindPath(type = 'best_tune_1' , sub_path = 'dataset/rossmann-store-sales/ensembles/best_tune_1',createDir = T) ## out 
+ff.bindPath(type = 'submission_1' , sub_path = 'dataset/rossmann-store-sales/ensembles/pred_ensemble_1',createDir = T) ## out 
+
+## DATA 
+cat("reading the train and test data\n")
+train <- read_csv(paste(ff.getPath("data") , "train.csv" , sep=''))
+test  <- read_csv(paste(ff.getPath("data") , "test.csv" , sep=''))
+store <- read_csv(paste(ff.getPath("data") , "store.csv" , sep=''))
+
+# removing the date column (since elements are extracted) and also StateHoliday which has a lot of NAs (may add it back in later)
+train <- merge(train,store)
+test <- merge(test,store)
+
+# There are some NAs in the integer columns so conversion to zero
+train[is.na(train)]   <- 0
+test[is.na(test)]   <- 0
+
+cat("train data column names and details\n")
+names(train)
+str(train)
+summary(train)
+cat("test data column names and details\n")
+names(test)
+str(test)
+summary(test)
+
+# looking at only stores that were open in the train set
+# may change this later
+train <- train[ which(train$Open=='1'),]
+train <- train[ which(train$Sales!='0'),]
+
+# seperating out the elements of the date column for the train set
+train$month <- as.integer(format(train$Date, "%m"))
+train$year <- as.integer(format(train$Date, "%y"))
+train$day <- as.integer(format(train$Date, "%d"))
+
+# removing the date column (since elements are extracted) and also StateHoliday which has a lot of NAs (may add it back in later)
+train <- train[,-c(3,8)]
+
+# seperating out the elements of the date column for the test set
+test$month <- as.integer(format(test$Date, "%m"))
+test$year <- as.integer(format(test$Date, "%y"))
+test$day <- as.integer(format(test$Date, "%d"))
+
+# removing the date column (since elements are extracted) and also StateHoliday which has a lot of NAs (may add it back in later)
+test <- test[,-c(4,7)]
+
+feature.names <- names(train)[c(1,2,5:19)]
+cat("Feature Names\n")
+feature.names
+
+
+cat("assuming text variables are categorical & replacing them with numeric ids\n")
+for (f in feature.names) {
+  if (class(train[[f]])=="character") {
+    levels <- unique(c(train[[f]], test[[f]]))
+    train[[f]] <- as.integer(factor(train[[f]], levels=levels))
+    test[[f]]  <- as.integer(factor(test[[f]],  levels=levels))
+  }
+}
+
+cat("train data column names after slight feature engineering\n")
+names(train)
+cat("test data column names after slight feature engineering\n")
+names(test)
+tra<-train[,feature.names]
+RMPSE<- function(preds, dtrain) {
+  labels <- getinfo(dtrain, "label")
+  elab<-exp(as.numeric(labels))-1
+  epreds<-exp(as.numeric(preds))-1
+  err <- sqrt(mean((epreds/elab-1)^2))
+  return(list(metric = "RMPSE", value = err))
+}
+nrow(train)
+h<-sample(nrow(train),10000)
+
+### exporting to data 
+train_export <- tra 
+train_export$Sales <- train$Sales
+test_export <- test[,feature.names]
+test_export$Id <- test$Id
+
+write_csv(train_export, paste(ff.getPath("elab") , "Xtrain_bench.csv" , sep=''))
+write_csv(test_export, paste(ff.getPath("elab") , "Xtest_bench.csv" , sep=''))
+### end exporting to data 
+
+##
+dval<-xgb.DMatrix(data=data.matrix(tra[h,]),label=log(train$Sales+1)[h])
+dtrain<-xgb.DMatrix(data=data.matrix(tra[-h,]),label=log(train$Sales+1)[-h])
+watchlist<-list(val=dval,train=dtrain)
+param <- list(  objective           = "reg:linear", 
+                booster = "gbtree",
+                eta                 = 0.005, # 0.06, #0.01,
+                max_depth           = 10, #changed from default of 8
+                subsample           = 0.9, # 0.7
+                colsample_bytree    = 0.7 # 0.7
+                #num_parallel_tree   = 2
+                # alpha = 0.0001, 
+                # lambda = 1
+)
+cat("****** param ***** \n")
+print(param)
+#### 
+clf <- xgb.train(   params              = param, 
+                    data                = dtrain, 
+                    nrounds             = 14000, #300, #280, #125, #250, # changed from 300
+                    verbose             = 0,
+                    early.stop.round    = 100,
+                    watchlist           = watchlist,
+                    maximize            = FALSE,
+                    feval=RMPSE
+)
+pred1 <- exp(predict(clf, data.matrix(test[,feature.names]))) -1
+submission <- data.frame(Id=test$Id, Sales=pred1)
+cat("saving the submission file\n")
+write_csv(submission, paste(ff.getPath("elab") , "bech_eta0005_nround14000.csv" , sep=''))
+
+
+
+
+
+
+
+
+
+
+
+