Getting Started with Pandas

A Minimal, Practical Guide for Python Developers

---

Pandas is the go-to Python library for working with structured, tabular data — think spreadsheets, CSVs, or SQL tables, but in code. It sits at the heart of most data science and machine learning workflows.

---

Installation

pip install pandas

---

Core Data Structures

Pandas has two primary building blocks:

Structure	Dimensions	Analogy
Series	1D	A single column
DataFrame	2D	A full table

Series

import pandas as pd

s = pd.Series([10, 20, 30], index=['a', 'b', 'c'])
print(s['b'])  # 20

DataFrame

df = pd.DataFrame({
    'name':  ['Alice', 'Bob', 'Charlie'],
    'age':   [25, 30, 35],
    'score': [88.5, 92.0, 79.5]
})

---

Selecting Data

df['name']            # single column → Series
df[['name', 'age']]  # multiple columns → DataFrame

df.loc[0]             # row by label / index
df.iloc[1]            # row by integer position

df[df['age'] > 28]    # filter rows by condition

---

Exploring a DataFrame

df.head(2)       # first N rows
df.tail(2)       # last N rows
df.shape         # (rows, cols)
df.dtypes        # column data types
df.describe()    # statistical summary

---

Common Operations

df['age'].mean()
df['score'].max()

# Group and aggregate
df.groupby('name')['score'].mean()

---

Adding & Modifying Columns

# Derive a new column
df['grade'] = df['score'].apply(lambda x: 'A' if x >= 90 else 'B')

# Modify in place
df['age'] = df['age'] + 1

---

Handling Missing Data

df.isnull().sum()   # count NaNs per column
df.dropna()         # drop rows containing any NaN
df.fillna(0)        # replace NaN with a default value

---

Gotcha: Chained Assignment

df[df['age'] > 28]['score'] = 100   # may not do what you expect

This selects a subset first, then assigns into it. Pandas cannot always tell whether that subset is a view or a copy, so the write may land on a temporary object and never reach df. You will usually see a SettingWithCopyWarning when this happens.

Do it in one step with .loc instead:

df.loc[df['age'] > 28, 'score'] = 100

The same rule applies when you slice a DataFrame and keep working on the result — call .copy() if you want the slice to be independent of its parent.

---

Reading & Writing Files

# Reading
df = pd.read_csv('data.csv')
df = pd.read_json('data.json')
df = pd.read_excel('data.xlsx')

# Writing
df.to_csv('output.csv', index=False)

---

Sorting & Merging

# Sort
df.sort_values('score', ascending=False)

# SQL-style JOIN
merged = pd.merge(df1, df2, on='id', how='inner')  # also: left, right, outer

# Stack rows
combined = pd.concat([df1, df2], ignore_index=True)

---

Mental Model (coming from C / plain Python)

C / Python	Pandas Equivalent
array[i]	df.iloc[i]
struct.field	df['column']
for loop filter	df[condition]
manual stats	df.describe()

---

Key Takeaway

Think column-first and vectorized. Avoid Python for loops over rows — instead, let pandas operate on entire columns at once. It's faster, more readable, and more idiomatic.

---

Next step: pair pandas with NumPy for numerical computing, or feed your cleaned DataFrames into PyTorch / scikit-learn for machine learning.