01. What is EDA & why it matters 02. Understand the problem 03. Import & inspect data 04. Handle missing values 05. Clean duplicates & types 06. Univariate analysis 07. Bivariate & correlation 08. Outlier detection 09. Feature relationships 10. Summarize & document 11. Tools for EDA 12. Common mistakes

01

Understand the problem before you touch the data

Every dataset exists to answer a question. Are you predicting customer churn, diagnosing a machine fault, or measuring the impact of a marketing campaign? Write that question down in one sentence before opening your notebook.

This step also means learning where the data came from — surveys, sensors, transaction logs, web forms — because the source tells you what kind of noise and bias to expect later. A field collected by a rushed call-centre agent behaves very differently from one logged automatically by a machine.

02

Import and get a first look at the data

Load the dataset and check its shape, column names, and data types before anything else. This single step catches a surprising number of problems — a numeric column stored as text, a date field that didn't parse correctly, or a file that loaded with the wrong number of rows.

import pandas as pd

df = pd.read_csv("sales_data.csv")

print(df.shape)        # rows, columns
print(df.dtypes)       # data type of each column
df.head()               # first 5 rows
df.info()                # nulls + memory usage in one view

03

Find and handle missing values

Missing data is rarely random. A blank "income" field might mean the customer skipped a sensitive question, or a null "delivery_date" might mean the order was cancelled. Before filling anything in, measure how much is missing and look for a pattern in where it's missing.

df.isnull().sum()
df.isnull().mean() * 100   # percentage missing per column

# Common strategies
df["age"].fillna(df["age"].median(), inplace=True)
df.dropna(subset=["customer_id"], inplace=True)

As a rule of thumb: a column missing more than ~40–50% of its values is often better dropped than imputed, unless that column is critical to your question.

04

Remove duplicates and fix data types

Duplicate rows quietly inflate counts and skew averages. Mismatched data types — like a price stored as a string with a currency symbol — silently break every calculation downstream. Clean both before you write a single chart.

df.duplicated().sum()
df.drop_duplicates(inplace=True)

df["price"] = df["price"].str.replace("₹", "").astype(float)
df["order_date"] = pd.to_datetime(df["order_date"])

05

Study each variable on its own — univariate analysis

Now look at one column at a time. For numeric columns, check the mean, median, and spread. For categorical columns, check the value counts. This is where you spot skewed distributions, rare categories, and values that don't belong — like an "age" of 250.

df.describe()                        # numeric summary
df["category"].value_counts()        # category frequencies

import seaborn as sns
sns.histplot(df["price"], kde=True)
sns.boxplot(x=df["category"])

06

Compare variables together — bivariate & correlation analysis

Once you understand each column individually, look at how pairs of columns relate. Does spend increase with age? Does churn rate change by region? A correlation matrix and a heatmap are the fastest way to scan dozens of relationships at once.

corr = df.corr(numeric_only=True)
sns.heatmap(corr, annot=True, cmap="coolwarm")

sns.scatterplot(x="marketing_spend", y="revenue", data=df)

A word of caution: correlation tells you two variables move together, not that one causes the other. Ice cream sales and drowning incidents both rise in summer — the real driver is heat, not each other.

07

Hunt down outliers

Outliers can be genuine (a rare million-dollar transaction) or errors (a negative age). The Interquartile Range (IQR) method is a simple, reliable way to flag them before deciding whether to cap, remove, or keep them as-is.

Q1 = df["revenue"].quantile(0.25)
Q3 = df["revenue"].quantile(0.75)
IQR = Q3 - Q1

lower = Q1 - 1.5 * IQR
upper = Q3 + 1.5 * IQR

outliers = df[(df["revenue"] < lower) | (df["revenue"] > upper)]

08

Visualize feature relationships for the full picture

With the individual pieces understood, bring them together. Pair plots, grouped bar charts, and time-series line plots often reveal patterns that no single number can — a seasonal spike, a segment that behaves differently, or a trend that reverses over time.

sns.pairplot(df[["age", "income", "spend", "churn"]], hue="churn")

df.groupby("region")["revenue"].mean().plot(kind="bar")

09

Summarize your findings and document decisions

Close the loop by writing down what you found: which columns were dropped and why, which values were imputed, which relationships look meaningful, and which questions remain open. This summary becomes the map that guides feature engineering and model selection — and it's what you'll thank yourself for six months from now when someone asks "why did we do it this way?"

"A model built on unexamined data is a guess dressed up in decimals. EDA is what turns that guess into something you can defend."