Exploratory data analysis in ML Python - Deep Dive

Overview - Exploratory data analysis

What is it?

Exploratory data analysis (EDA) is the process of examining and understanding data before building models. It involves summarizing the main characteristics, finding patterns, spotting anomalies, and checking assumptions using visual and statistical methods. EDA helps you get to know your data deeply and prepares it for further analysis or modeling. It is like getting to know a new friend by asking questions and observing carefully.

Why it matters

Without EDA, you might build models on data that has errors, missing values, or hidden patterns that mislead your results. EDA helps prevent costly mistakes by revealing the true nature of your data early. It saves time and improves model quality by guiding data cleaning, feature selection, and hypothesis formation. In real life, skipping EDA is like trying to fix a car without checking what’s wrong first.

Where it fits

Before EDA, you should know basic data types and how to load data into your tools. After EDA, you move on to data cleaning, feature engineering, and then model building. EDA is the bridge between raw data and machine learning models.

Mental Model

Core Idea

Exploratory data analysis is like detective work that uncovers the hidden story and quirks in your data before you make decisions.

Think of it like...

Imagine you just got a box of assorted fruits from a market. Before making a fruit salad, you look at each fruit, smell it, check for bruises, and sort them by type and ripeness. This helps you decide which fruits to use and how to prepare them. EDA is the same but with data.

┌─────────────────────────────┐
│        Raw Data              │
└─────────────┬───────────────┘
              │
              ▼
┌─────────────────────────────┐
│ Exploratory Data Analysis    │
│ - Summary stats             │
│ - Visualizations            │
│ - Detect anomalies          │
│ - Understand distributions  │
└─────────────┬───────────────┘
              │
              ▼
┌─────────────────────────────┐
│ Cleaned & Understood Data   │
│ Ready for Modeling          │
└─────────────────────────────┘

Build-Up - 7 Steps

1

Foundation

What is Exploratory Data Analysis

Concept: Introducing the basic idea of EDA as the first step to understand data.

Exploratory data analysis means looking at your data carefully to learn about it. You check how many rows and columns it has, what types of data are inside, and get simple summaries like averages or counts. This helps you know what you are working with before doing anything complex.

Result

You get a basic understanding of your dataset’s size, types, and simple statistics.

Understanding your data’s shape and type is the foundation for all further analysis and prevents surprises later.

2

Foundation

Summary Statistics Basics

Concept: Learn how to calculate and interpret simple statistics like mean, median, and counts.

Summary statistics give you quick numbers that describe your data. For example, the mean tells you the average value, the median shows the middle value, and counts tell you how many times each category appears. These numbers help you see the center, spread, and balance of your data.

Result

You can describe your data with numbers that summarize its main features.

Knowing summary statistics helps you spot if data is skewed, has outliers, or is balanced.

3

Intermediate

Visualizing Data Distributions

🤔Before reading on: do you think a histogram shows individual data points or the overall shape of data? Commit to your answer.

Concept: Using charts like histograms and boxplots to see how data values spread and where they cluster.

Histograms group data into bins and show how many values fall into each bin, revealing the shape of the data distribution. Boxplots show the median, quartiles, and outliers visually. These plots help you understand if data is normal, skewed, or has unusual values.

Result

You can visually identify patterns like skewness, gaps, or outliers in your data.

Visual tools reveal data stories that numbers alone can hide, making patterns and problems easier to spot.

4

Intermediate

Detecting Missing and Anomalous Data

🤔Before reading on: do you think missing data always means a mistake, or can it sometimes be meaningful? Commit to your answer.

Concept: Learn to find missing values and unusual data points that may affect analysis.

Missing data can appear as blanks or special markers. Anomalies are values that don’t fit the usual pattern, like a height of 300 cm. Detecting these helps you decide whether to fix, remove, or keep them. Sometimes missing data tells a story, like a skipped survey question.

Result

You identify data quality issues that need attention before modeling.

Recognizing missing and strange data early prevents errors and improves model trustworthiness.

5

Intermediate

Exploring Relationships Between Variables

🤔Before reading on: do you think two variables with a strong relationship always cause each other? Commit to your answer.

Concept: Using scatter plots and correlation to see how variables relate or move together.

Scatter plots show points for pairs of variables, revealing trends or clusters. Correlation measures how strongly two variables move together, from -1 (opposite) to +1 (same direction). This helps find useful connections or redundant data.

Result

You understand which variables might influence each other or the target.

Knowing variable relationships guides feature selection and hypothesis building.

6

Advanced

Using EDA to Guide Feature Engineering

🤔Before reading on: do you think EDA only helps with cleaning data, or can it also inspire new features? Commit to your answer.

Concept: Applying insights from EDA to create or transform variables that improve models.

By understanding distributions and relationships, you can create new features like combining variables, encoding categories, or transforming skewed data. For example, if age and income relate to buying behavior, you might create an age-income ratio feature.

Result

Your data becomes richer and more informative for machine learning.

EDA is not just about cleaning but also about discovering new ways to represent data.

7

Expert

Pitfalls and Surprises in EDA Interpretation

🤔Before reading on: do you think all patterns found in EDA are reliable and useful? Commit to your answer.

Concept: Understanding that EDA can mislead if not done carefully, and how to avoid common traps.

Patterns in EDA might be due to random chance, sampling bias, or data leakage. For example, a strong correlation might disappear in new data. Experts use statistical tests, cross-validation, and domain knowledge to confirm findings. Also, visualizations can exaggerate or hide details depending on scale and binning.

Result

You develop a cautious and critical mindset when interpreting EDA results.

Knowing EDA’s limits prevents overconfidence and guides better decisions in modeling.

Under the Hood

EDA works by applying simple mathematical summaries and visual mappings to raw data arrays or tables. Internally, it calculates statistics like sums, means, and counts by iterating over data points. Visualizations transform numeric data into graphical elements like bars or points using coordinate systems. Missing data detection scans for null or special values. Correlation computes covariance normalized by variance. These operations are efficient and often vectorized in software libraries.

Why designed this way?

EDA was designed to give humans an intuitive and quick way to understand complex data without building full models. Early statisticians like John Tukey emphasized visual and simple numeric summaries to explore data patterns before formal analysis. This approach balances speed, interpretability, and insight, avoiding premature assumptions. Alternatives like jumping straight to modeling risk missing data issues or hidden patterns.

┌───────────────┐
│   Raw Data    │
└──────┬────────┘
       │
       ▼
┌───────────────┐
│ Compute Stats │
│ (mean, median)│
└──────┬────────┘
       │
       ▼
┌───────────────┐       ┌───────────────┐
│ Visualizations│◄─────►│ Missing Data  │
│ (histogram,   │       │ Detection     │
│  boxplot)     │       └───────────────┘
└──────┬────────┘
       │
       ▼
┌───────────────┐
│ Correlations  │
│ & Relationships│
└──────┬────────┘
       │
       ▼
┌───────────────┐
│ Insights for  │
│ Cleaning &    │
│ Feature Eng.  │
└───────────────┘

Myth Busters - 4 Common Misconceptions

Quick: Does a high correlation between two variables mean one causes the other? Commit to yes or no.

Common Belief: High correlation means one variable causes the other.

Reality: Correlation only shows association, not causation. Two variables can move together due to a third factor or coincidence.

Why it matters: Mistaking correlation for causation can lead to wrong conclusions and poor decisions in modeling and business.

Quick: Is missing data always a mistake that should be deleted? Commit to yes or no.

Common Belief: Missing data is always an error and should be removed.

Reality: Missing data can be meaningful, indicating absence or special cases. Removing it blindly can bias results.

Why it matters: Ignoring the meaning of missing data can distort analysis and reduce model accuracy.

Quick: Does a beautiful visualization always mean the data is reliable? Commit to yes or no.

Common Belief: If a visualization looks clear and neat, the data must be good.

Reality: Visualizations can be misleading due to scale choices, bin sizes, or selective data. They need critical interpretation.

Why it matters: Relying blindly on visuals can hide problems and lead to wrong insights.

Quick: Does EDA guarantee you will find the best features for your model? Commit to yes or no.

Common Belief: EDA automatically reveals the best features for modeling.

Reality: EDA helps suggest features but does not guarantee their predictive power. Model testing is needed.

Why it matters: Overtrusting EDA can waste time on irrelevant features and hurt model performance.

Expert Zone

1

EDA results depend heavily on the sample of data; small or biased samples can mislead even expert analysis.

2

The choice of visualization parameters like bin size or axis scale can drastically change the story the data tells.

3

Combining domain knowledge with EDA insights is crucial; pure statistics without context often misses key patterns.

When NOT to use

EDA is less useful when working with extremely large streaming data where real-time summaries are needed; instead, incremental or automated monitoring tools are better. Also, for fully synthetic or simulated data with known properties, EDA may be redundant.

Production Patterns

In real-world projects, EDA is integrated into automated pipelines with reports and dashboards. Teams use EDA to validate new data batches, monitor data drift, and guide feature engineering cycles. EDA outputs often feed into data versioning and model explainability tools.

Connections

Data Cleaning

Builds-on

Understanding data issues through EDA directly informs how to clean and prepare data effectively.

Statistical Hypothesis Testing

Builds-on

EDA helps form hypotheses about data patterns that can later be tested rigorously with statistics.

Journalism

Similar pattern

Like journalists investigate facts and stories before writing, EDA investigates data to uncover its story before modeling.

Common Pitfalls

#1 Ignoring missing data or treating it all the same way.

Wrong approach:

data = data.dropna() # Remove all missing values without checking

Correct approach:

missing_summary = data.isnull().sum()
# Analyze missing patterns before deciding how to handle

Root cause:

Assuming missing data is always an error leads to careless removal that can bias results.

#2 Using inappropriate visualization scales that hide data details.

Wrong approach:

plt.hist(data['income'], bins=5) # Too few bins hides distribution shape

Correct approach:

plt.hist(data['income'], bins=30) # More bins reveal detailed distribution

Root cause:

Not tuning visualization parameters causes misleading or oversimplified views.

#3 Assuming correlation implies causation.

Wrong approach:

print(data['A'].corr(data['B']))
# Conclude A causes B without further analysis

Correct approach:

# Use domain knowledge and experiments to test causation beyond correlation

Root cause:

Misunderstanding statistical association leads to wrong causal claims.

Key Takeaways

Exploratory data analysis is the essential first step to understand and prepare data before modeling.

Using summary statistics and visualizations reveals patterns, anomalies, and relationships in data.

Detecting missing and unusual data early prevents errors and improves model quality.

EDA insights guide feature engineering and cleaning, but require critical interpretation to avoid pitfalls.

Expert EDA balances automated tools with domain knowledge and cautious skepticism about patterns.

Practice

1. What is the main goal of exploratory data analysis (EDA) in machine learning?

easy

A. To train the machine learning model

B. To understand the main features and quality of the data

C. To deploy the model to production

D. To write the final report without looking at data

Solution

1. Step 1: Understand the purpose of EDA

   EDA is used to explore and understand the data before modeling.

2. Step 2: Differentiate EDA from other steps

   Training and deployment happen after EDA, so they are not the main goal of EDA.

3. Final Answer:

   To understand the main features and quality of the data -> Option B

4. Quick Check:

   EDA = Understand data [OK]

Hint: EDA means exploring data, not modeling or deploying [OK]

Common Mistakes:

• Confusing EDA with model training
• Thinking EDA is for deployment
• Skipping data understanding step

2. Which Python library is commonly used for basic exploratory data analysis tasks like viewing data samples and summary statistics?

easy

A. matplotlib

B. tensorflow

C. pandas

D. scikit-learn

Solution

1. Step 1: Identify libraries for EDA

   Pandas provides functions like head() and describe() for data samples and summary stats.

2. Step 2: Exclude other libraries

   Matplotlib is for plotting, scikit-learn for modeling, tensorflow for deep learning, not basic EDA.

3. Final Answer:

   pandas -> Option C

4. Quick Check:

   pandas = data samples and stats [OK]

Hint: Use pandas for quick data overview and stats [OK]

Common Mistakes:

• Using matplotlib for data samples
• Confusing modeling libraries with EDA
• Trying tensorflow for simple stats

3. Given the following Python code using pandas, what will print(df.describe()) output?

import pandas as pd
data = {'age': [25, 30, 22, 40, 28], 'score': [88, 92, 85, 95, 90]}
df = pd.DataFrame(data)
print(df.describe())

medium

A. Summary statistics including count, mean, std, min, max for 'age' and 'score'

B. The first 5 rows of the dataframe

C. A plot showing distribution of 'age' and 'score'

D. An error because describe() needs parameters

Solution

1. Step 1: Understand what describe() does

   It returns summary statistics like count, mean, std, min, max for numeric columns.

2. Step 2: Check the code output

   Since 'age' and 'score' are numeric, describe() will show stats for both columns.

3. Final Answer:

   Summary statistics including count, mean, std, min, max for 'age' and 'score' -> Option A

4. Quick Check:

   describe() = summary stats [OK]

Hint: describe() shows stats, head() shows rows [OK]

Common Mistakes:

• Thinking describe() shows data rows
• Expecting a plot from describe()
• Assuming describe() needs arguments

4. You run this code but get an error:

import pandas as pd
data = {'age': [25, 30, None, 40], 'score': [88, 92, 85, 95]}
df = pd.DataFrame(data)
print(df.mean())

What is the likely cause and fix?

medium

A. Error due to None value; fix by using df.mean(skipna=True)

B. Error due to missing import; fix by importing numpy

C. Error due to wrong data type; fix by converting all to strings

D. No error; code runs fine as is

Solution

1. Step 1: Check how pandas handles None in numeric columns

   Pandas treats None as NaN and by default mean() skips NaN values.

2. Step 2: Verify if code runs without error

   The code runs fine and calculates mean ignoring None.

3. Final Answer:

   No error; code runs fine as is -> Option D

4. Quick Check:

   pandas mean() skips None by default [OK]

Hint: pandas mean() ignores None/NaN by default [OK]

Common Mistakes:

• Assuming None causes error in mean()
• Thinking numpy import is needed here
• Trying to convert numbers to strings

5. You have a dataset with columns age, income, and purchased (Yes/No). Which EDA step helps you find if income affects purchased?

hard

A. Plotting boxplots of income grouped by purchased categories

B. Calculating mean age of all customers

C. Checking for missing values in age column

D. Sorting data by income in ascending order

Solution

1. Step 1: Identify relationship to explore

   We want to see if income differs between those who purchased and those who did not.

2. Step 2: Choose appropriate EDA method

   Boxplots grouped by purchased categories show income distribution differences clearly.

3. Final Answer:

   Plotting boxplots of income grouped by purchased categories -> Option A

4. Quick Check:

   Boxplots show group differences [OK]

Hint: Use grouped boxplots to compare numeric by category [OK]

Common Mistakes:

• Ignoring grouping by purchased
• Checking unrelated columns
• Just sorting without visualization