R ile Makine Öğrenmesi: Veri Hazırlığı ve Keşif

Most machine learning articles start with the model. Yet in a real project most of the time is spent before any model exists: getting to know the data, finding what is missing, seeing how the variables are distributed, and splitting the data into training and test sets.

This is the first of a four-part series. Throughout the series we work with two real data sets: Carseats (child car seat sales at 400 stores) and the World Happiness Report 2021 (happiness scores for 149 countries). In this first post I cover everything that has to happen before a single model is fitted.

“If you do not know what the data is before you fit a model, what you built is not a model — it is a guess.”

1. Two Data Sets

Two data sets run through the whole series. Both are open and both are small, so you can reproduce every result on your own machine.

library(tidyverse)
library(ISLR)

data(Carseats)
dim(Carseats)      # 400  11
str(Carseats)
summary(Carseats$Sales)

   Min. 1st Qu.  Median    Mean 3rd Qu.    Max.
  0.000   5.390   7.490   7.496   9.320  16.270

The mean and the median being this close (7.496 and 7.49) is the first clue: Sales is close to symmetric. We will confirm that with a number shortly.

2. Missing Data Analysis

Checking for missing data is the first step that must not be skipped before modelling. R offers several practical ways to do it.

# Quickest: total and per column
sum(is.na(Carseats))                 # 0
colSums(is.na(Carseats))

# Visually
library(VIM)
aggr(Carseats, numbers = TRUE, sortVars = TRUE)

library(Amelia)
missmap(Carseats)

Carseats has no missing values at all — expected for a clean textbook data set. Real projects are rarely like this, which is why why a value is missing matters as much as how you handle it.

3. Exploring the Data

The distribution of the categorical variables matters as much as the numeric summary. An imbalanced category affects both what the model can learn and the stratified split you will do later.

table(Carseats$ShelveLoc)
table(Carseats$Urban)
table(Carseats$US)

How are the variables related to each other?

Individual distributions matter, but so does how the variables sit relative to one another. A correlation matrix shows which variables are closest to the target and warns early about dependencies that could cause trouble in the model.

numeric_cols <- Carseats[sapply(Carseats, is.numeric)]
co <- cor(numeric_cols)

round(sort(co[, "Sales"], decreasing = TRUE), 3)

# visualised with corrplot:
library(corrplot)
corrplot(co, method = "color", type = "upper", addCoef.col = "black")

Not one variable clears 0.5; the strongest relationship is with Price and even that is −0.445. This is the first sign that a univariate view is not enough. In the regression post we will see these same variables explain 87% of sales together — the information lives in the combination, not in any single column.

The relationships among the predictors deserve a look as well:

Outliers: are there any, and should they go?

Outlier analysis is often applied as "find them and delete them", which is usually wrong. The first step is simply to see how many there are.

for (v in names(numeric_cols)) {
  x <- numeric_cols[[v]]
  q <- quantile(x, c(.25, .75)); i <- IQR(x)
  cat(v, sum(x < q[1] - 1.5*i | x > q[2] + 1.5*i), "\n")
}

Nine outliers across 400 observations — 2.25%. None of them is physically impossible either: a store selling 16,270 units is plausible, and so is a product priced above 177. These are not data errors, they are the tails of the distribution.

Which chart answers which question is a topic I covered in a separate post; instead of repeating it, here is a short mapping.

4. Normality, Skewness and Kurtosis

The assumption in linear regression is that the error term is normally distributed, not the variables themselves. Still, knowing the distributions up front tells you which transformation might be needed and which chart could mislead you.

skewness <- function(x) mean((x - mean(x))^3) / sd(x)^3
kurtosis <- function(x) mean((x - mean(x))^4) / sd(x)^4

numeric_vars <- Carseats[sapply(Carseats, is.numeric)]

for (v in names(numeric_vars)) {
  sw <- shapiro.test(numeric_vars[[v]])
  cat(sprintf("%-12s skew=%6.3f kurt=%6.3f W=%.4f p=%.4g\n",
              v, skewness(numeric_vars[[v]]), kurtosis(numeric_vars[[v]]),
              sw$statistic, sw$p.value))
}

One thing stands out: although the skewness values are very close to zero, normality is rejected for five variables. The reason is kurtosis. For Income, Population, Age and Education it sits between 1.7 and 1.9, clearly below the value of 3 for a normal distribution. These variables are symmetric but far too flat — close to a uniform distribution. Symmetry alone does not mean normality.

5. Splitting into Training and Test Sets

If you measure a model with the data it was trained on, the result always looks better than it is — the model has already seen those rows. Real performance is measured on data the model has never seen.

For classification a plain random split may not be enough. If the class ratio is not preserved in both sets, comparisons break down. The fix is a stratified split: sampling the same proportion from each class.

df <- Carseats
df$Sales <- factor(ifelse(df$Sales <= 8, "Low", "High"),
                   levels = c("Low", "High"))

set.seed(5364)
idx <- unlist(lapply(levels(df$Sales), function(lv) {
  rows <- which(df$Sales == lv)
  sample(rows, size = floor(0.8 * length(rows)))
}))

train <- df[idx, ]
test  <- df[-idx, ]

table(train$Sales)   # Low 188  High 131
table(test$Sales)    # Low  48  High  33

createDataPartition() from the caret package does the same job; the explicit version above shows what happens underneath. The result: 319 training and 81 test observations, with roughly 59% Low / 41% High in both.

Is putting the threshold at 8 an innocent decision?

For classification we split Sales in two at 8,000 units. That number does not come from the data, it comes from us — and it sets the base rate directly.

Had we cut at 7.5, the base rate would be 50.5% and "60% accuracy" would count as a genuine result. Cut at 9 and the base rate rises to 71.8%, where the same accuracy means failure. An accuracy figure has to be read together with the threshold that produced it.

There is a second issue: the stores near the line. 56 stores have Sales between 7.5 and 8.5 — 14% of the dataset. Nothing meaningful separates them into "High" and "Low"; only which side of the line they happen to land on.

6. Avoiding Data Leakage

Data leakage is information from the test set reaching the model unnoticed. The outcome is always the same: the model looks excellent in the report and mediocre in production.

The classification example in this series carried a hidden leakage risk: when Sales is turned into a binary variable, the original numeric column must be overwritten. Had numeric Sales stayed in the data, the model would read the target directly and report 100% accuracy.

7. A Checklist Before Modelling

Conclusion

We did not fit a single model in this post, but we laid the whole foundation for the next step: we got to know the data, confirmed there are no missing values, measured how the variables are distributed, and split the data reproducibly into training and test sets.

Two facts from here will matter directly in the next posts: Sales, CompPrice and Price are consistent with a normal distribution, and the base rate in classification is roughly 59%. That second number is especially important — because a model reporting 60% accuracy may have learned nothing at all.

“A good model starts with well-prepared data; a good evaluation starts the moment you split it correctly.”