Load required libraries

# Suppress startup messages of library dplyr
suppressPackageStartupMessages(library(dplyr))
# Loading required libraries
library(dplyr, quietly = TRUE)
library(class, quietly = TRUE)

0. Load raw dataset

ifood <- read.csv("ml_project1_data.csv", sep=",", header=TRUE, stringsAsFactors = FALSE)

1. Remove irrelevant columns

ifood <- ifood[, !names(ifood) %in% c("ID", "Z_CostContact", "Z_Revenue")]

3. Rename columns for easier access

colnames(ifood) <- gsub("NumDealsPurchases", "DealsPurc", colnames(ifood))
colnames(ifood) <- gsub("NumWebPurchases", "WebPurc", colnames(ifood))
colnames(ifood) <- gsub("NumStorePurchases", "StorePurc", colnames(ifood))
colnames(ifood) <- gsub("NumWebVisitsMonth", "WebVisits", colnames(ifood))
colnames(ifood) <- gsub("AcceptedCmpOverall", "CmpOverall", colnames(ifood))
colnames(ifood) <- gsub("MntWines", "WineExp", colnames(ifood))
colnames(ifood) <- gsub("MntFruits", "FruitExp", colnames(ifood))
colnames(ifood) <- gsub("MntMeatProducts", "MeatExp", colnames(ifood))
colnames(ifood) <- gsub("MntFishProducts", "FishExp", colnames(ifood))
colnames(ifood) <- gsub("MntSweetProducts", "SweetExp", colnames(ifood))
colnames(ifood) <- gsub("MntGoldProds", "GoldExp", colnames(ifood))
colnames(ifood) <- gsub("Marital_Status", "MaritalSts", colnames(ifood))
colnames(ifood) <- gsub("NumCatalogPurchases", "CatalogPurc", colnames(ifood))
colnames(ifood) <- gsub("AcceptedCmp1", "AccCmp1", colnames(ifood))
colnames(ifood) <- gsub("AcceptedCmp2", "AccCmp2", colnames(ifood))
colnames(ifood) <- gsub("AcceptedCmp3", "AccCmp3", colnames(ifood))
colnames(ifood) <- gsub("AcceptedCmp4", "AccCmp4", colnames(ifood))
colnames(ifood) <- gsub("AcceptedCmp5", "AccCmp5", colnames(ifood))

4. Handle outliers

ifood$Age <- ifelse(ifood$Age > 80, 80, ifood$Age)

5. Handle missing values

ifood <- ifood[!ifood$MaritalSts %in% c("YOLO", "Absurd"),]
ifood$MaritalSts[ifood$MaritalSts == "Alone"] <- "Single"

6. Impute missing Income using KNN

ifood$Income <- ifelse(ifood$Income < 12500, NA, ifood$Income)

num_vars <- sapply(ifood, is.numeric)
complete_vars <- colnames(ifood)[num_vars]
missing_threshold <- 0.2 * nrow(ifood)
complete_vars <- complete_vars[colSums(is.na(ifood[, complete_vars])) < missing_threshold]
aux <- ifood[, complete_vars]

var <- "Income"
aux1 <- aux[!is.na(ifood[[var]]), , drop = FALSE]
aux2 <- aux[is.na(ifood[[var]]), , drop = FALSE]

cols_na <- colnames(aux2)[colSums(is.na(aux2)) > 0]
if (length(cols_na) > 0) {
  aux1 <- aux1[, !(colnames(aux1) %in% cols_na), drop = FALSE]
  aux2 <- aux2[, !(colnames(aux2) %in% cols_na), drop = FALSE]
}

knn_impute <- knn(aux1, aux2, ifood[[var]][!is.na(ifood[[var]])], k = 1)
ifood[[var]][is.na(ifood[[var]])] <- as.numeric(as.character(knn_impute))

7. Correct calculation of TotAccCmp

ifood$TotAccCmp <- ifood$AccCmp1 + ifood$AccCmp2 + ifood$AccCmp3 + ifood$AccCmp4 + ifood$AccCmp5

8. Remove duplicate records

ifood <- ifood %>% arrange(desc(Response)) %>% distinct_at(vars(-Response), .keep_all = TRUE)

9. Create TotalExp before using it

ifood$TotalExp <- rowSums(ifood[, c("WineExp", "FruitExp", "MeatExp", "FishExp", "SweetExp", "GoldExp")], na.rm = TRUE)

10. Save cleaned dataset

write.csv(ifood, "ifood_cleaned.csv", row.names = FALSE)

Variable Creation

Second-Generation

Total Purchases

ifood$TotalPurchases <- ifood$DealsPurc + ifood$WebPurc + ifood$CatalogPurc + ifood$StorePurc

Purchase Frequency

ifood$PurchaseFrequency <- ifelse(ifood$CustDays > 0, ifood$TotalPurchases / (ifood$CustDays / 30), 0)

Preferred Product Category

product_categories <- c("WineExp", "FruitExp", "MeatExp", "FishExp", "SweetExp", "GoldExp")
max_index <- apply(ifood[ , product_categories], 1, which.max)
ifood$PreferredProductCategory <- product_categories[max_index]
ifood$PreferredProductCategory <- as.factor(ifood$PreferredProductCategory)

Preferred Purchase Channel

channels <- c("DealsPurc", "WebPurc", "CatalogPurc", "StorePurc")
max_ch_index <- apply(ifood[ , channels], 1, which.max)
ifood$PreferredChannel <- channels[max_ch_index]
ifood$PreferredChannel <- as.factor(ifood$PreferredChannel)

Average Spend Per Purchase

ifood$AvgSpendPerPurchase <- ifelse(ifood$TotalPurchases > 0, ifood$TotalExp / ifood$TotalPurchases, 0)

HasChildren

ifood$HasChildren <- ifelse(ifood$Kidhome + ifood$Teenhome > 0, 1, 0)

IncomeSegment

income_quantiles <- quantile(ifood$Income, probs = c(0.33, 0.66), na.rm = TRUE)
ifood$IncomeSegment <- cut(ifood$Income, breaks = c(-Inf, income_quantiles[1], income_quantiles[2], Inf),
                           labels = c("Low", "Medium", "High"))

CustomerTenure

ifood$CustomerTenure <- ifood$CustDays / 365

CampaignAcceptanceRate

ifood$CampaignAcceptanceRate <- ifelse(ifood$TotAccCmp > 0, ifood$TotAccCmp / 5, 0)

Third-Generation

Third-Generation Feature 1: Customer Segmentation via Clustering

Prepare data for clustering: use Recency, TotalPurchases (frequency), and TotalExp (monetary)

cluster_data <- ifood %>% select(Recency, TotalPurchases, TotalExp)

Scale the data for clustering

cluster_data_scaled <- scale(cluster_data)

Perform k-means clustering with 3 clusters (as an example)

set.seed(123)  # for reproducibility
k3 <- kmeans(cluster_data_scaled, centers = 3, nstart = 25)  # nstart for better convergence

Add the cluster assignment as a new feature

ifood$CustomerSegment <- as.factor(k3$cluster)

(Customers are now labeled 1, 2, or 3 based on their cluster segment)

Third-Generation Feature 2: Propensity Score via Logistic Regression

Fit a logistic regression model to predict campaign response (Response) using relevant features

propensity_model <- glm(Response ~ Income + Recency + TotalExp + TotalPurchases + TotAccCmp + Age + MaritalSts,
                        data = ifood, family = binomial)

Get predicted probabilities (propensity to respond)

ifood$PropensityScore <- predict(propensity_model, ifood, type = "response")

(PropensityScore is the model’s predicted probability of Response=1 for each customer)

Quick summary of PropensityScore range

summary(ifood$PropensityScore)
##    Min. 1st Qu.  Median    Mean 3rd Qu.    Max. 
## 0.00000 0.03488 0.07725 0.15313 0.18046 0.99312

Third-Generation Feature 3: Engagement Index

Normalize components between 0 and 1

Note: For Recency, a lower value means more recent (more engaged), so we invert it.

recency_norm   <- (max(ifood$Recency) - ifood$Recency) / max(ifood$Recency)          # invert recency
frequency_norm <- ifood$TotalPurchases / max(ifood$TotalPurchases)                   # purchases normalized
monetary_norm  <- ifood$TotalExp / max(ifood$TotalExp)                               # spending normalized
campaign_norm  <- (ifood$TotAccCmp + ifood$Response) / 6                             # campaign acceptance (out of 6 campaigns total including last response)
webvisit_norm  <- ifood$WebVisits / max(ifood$WebVisits)                             # web visits normalized

Calculate engagement index as average of all five components, scaled to 0-100

ifood$EngagementIndex <- (recency_norm + frequency_norm + monetary_norm + campaign_norm + webvisit_norm) / 5 * 100

Preview EngagementIndex distribution

summary(ifood$EngagementIndex)
##    Min. 1st Qu.  Median    Mean 3rd Qu.    Max. 
##   2.892  20.859  27.401  28.453  35.209  62.573

Save enriched dataset

write.csv(ifood, "ifood_enriched.csv", row.names = FALSE)