Data Visualization and Exploratory Data Analysis

Introduction

Business analytics is not only concerned with calculating statistics; it is concerned with discovering what data may be saying and communicating that meaning responsibly. Data visualization converts observations into graphical forms that can reveal comparisons, trends, distributions, relationships and unusual cases. Exploratory data analysis (EDA) is the systematic process of examining data before formal modelling or decision-making. It combines summaries, visual displays, data-quality checks and business reasoning.

EDA is especially important because business datasets are rarely perfect. They may contain missing values, duplicated records, inconsistent categories, extreme observations, measurement errors or misleading scales. A sound analyst therefore asks not only “What is the result?” but also “How were the data collected?”, “What assumptions are being made?” and “Could this presentation lead decision-makers to the wrong conclusion?”

This chapter develops practical foundations for tables, charts, dashboards and exploratory techniques. It also explains how to select appropriate displays, identify patterns and anomalies, examine relationships, and communicate findings ethically.

Learning Objectives

By the end of this chapter, you should be able to:

1. Distinguish categorical, ordinal, discrete numerical and continuous numerical data.
2. Construct and interpret frequency tables, cross-tabulations and appropriate charts.
3. Summarise distributions using location, variation and shape measures.
4. Use EDA to investigate missing values, outliers, trends and relationships.
5. Explain correlation and distinguish association from causation.
6. Design dashboards that support focused business decisions.
7. Recognise misleading visual presentations and communicate uncertainty clearly.
8. Apply an ethical, reproducible workflow to a business dataset.

1. The Role of Visualization in Business Analytics

A visualization is a deliberate mapping from data values to visual properties such as position, length, colour, shape or size. Effective displays reduce the mental effort required to compare values. For example, a bar chart usually makes category comparisons easier than a paragraph of figures, while a line chart makes a time trend easier to see than separate monthly values.

Visualization serves several purposes:

• Description: showing what has happened, such as quarterly revenue.
• Comparison: identifying which branches, products or customer groups differ.
• Distribution analysis: showing concentration, spread and unusual values.
• Relationship analysis: examining whether two variables move together.
• Monitoring: signalling whether performance is within an expected range.
• Communication: presenting evidence to managers and other stakeholders.

Visualization does not replace reasoning. A graph can show that two measures move together, but it cannot by itself establish why they move together. The analyst must combine the display with knowledge of definitions, measurement processes, sampling and business context.

2. Data Types and the Grammar of Displays

The type of variable determines which operations and visual encodings are appropriate.

A variable’s measurement level affects interpretation. It is not generally meaningful to calculate an average of payment methods coded 1, 2 and 3. Conversely, treating a numerical measurement such as delivery time as a set of unrelated categories may conceal useful distributional information.

A dataset can also be described by its observational unit, such as a customer, transaction, shipment or branch. Analysts must confirm the unit before counting records. Ten thousand transaction rows do not necessarily represent ten thousand customers.

3. Tables for Organising Data

Frequency tables

A frequency table counts observations in each category or class. If category i occurs fᵢ times in a dataset of n observations, its relative frequency is:

Here, fᵢ is the frequency for category i, n is the total number of valid observations, and pᵢ is a proportion. Multiplying pᵢ by 100 gives a percentage. The formula assumes that categories are mutually exclusive and collectively exhaustive for the observations being counted.

For continuous values, class intervals should be clearly defined. Intervals such as 0–10 and 10–20 can create ambiguity if the boundary value 10 is not assigned explicitly. A convention such as 0 ≤ x < 10 and 10 ≤ x < 20 avoids double counting.

Cross-tabulations

A cross-tabulation, or contingency table, displays counts for combinations of two categorical variables. For example, a retailer may compare purchase channel with returned/not returned status. Row percentages answer questions within each row; column percentages answer questions within each column. These percentages must not be confused.

If a cell contains nᵢⱼ observations, the row proportion is:

where nᵢ· is the total for row i. The column proportion is:

where n·ⱼ is the total for column j. The choice depends on the business question.

Example: interpreting percentages

Suppose 120 online orders were recorded: 30 were returned. Of 60 orders placed by new customers, 24 were returned; of 60 orders placed by existing customers, 6 were returned. The return rates are:

• New customers: 24 ÷ 60 = 0.40, or 40%.
• Existing customers: 6 ÷ 60 = 0.10, or 10%.

The conclusion is that the observed return rate for new customers is higher. It would be incorrect to say that 80% of all returns came from new customers without also explaining that 24 of the 30 returns came from that group.

4. Selecting and Designing Charts

Bar charts

Bar charts are appropriate for comparing separate categories. Bars should normally have a common baseline, equal widths and clearly labelled categories. Sorting categories by value often improves comparison, unless the natural order is meaningful.

Histograms

A histogram displays a numerical distribution using adjacent intervals. Unlike a bar chart, its bars touch because the intervals represent a continuous scale. The choice of bin width can materially change the appearance. Very narrow bins may create distracting noise; very wide bins may hide multiple peaks.

Box plots

A box plot summarises a distribution using the minimum or lower whisker, first quartile, median, third quartile and maximum or upper whisker, depending on the convention. The interquartile range is:

where Q₁ is the first quartile and Q₃ is the third quartile. A common rule flags observations below Q₁ − 1.5 × IQR or above Q₃ + 1.5 × IQR as potential outliers. This is a screening rule, not proof that an observation is erroneous. It is most useful for distributions where quartiles provide a meaningful summary.

Line charts

Line charts are suitable for ordered time observations or another naturally ordered numerical variable. Points should be connected only when the connection communicates continuity or sequence. Irregularly spaced dates should be plotted using their actual time spacing where possible.

Scatterplots

A scatterplot shows paired numerical observations. It can reveal direction, form, strength, clusters and unusual points. A fitted trend line may assist description, but it does not establish causation.

Dot plots, heat maps and maps

Dot plots can show individual values for small datasets. Heat maps use colour intensity to display a matrix, such as sales by product and region; their legend must explain whether darker colour means larger or smaller values. Geographic maps are appropriate when location is central to the question, but area and colour can distort perceived magnitude. A table may be better when precise comparison is required.

5. Summarising Numerical Distributions

The arithmetic mean is:

where xᵢ is observation i, n is the number of observations and x̄ is the sample mean. It uses every value but is sensitive to extreme observations.

The median is the middle ordered value, or the average of the two middle values when n is even. It is less affected by extreme values. The mode is the most frequent value or category.

The sample variance is:

and the sample standard deviation is:

The denominator n − 1 is used when estimating population variability from a sample under standard statistical conditions. If the data constitute the entire population of interest, a population variance may instead use denominator N. Analysts should state which convention is being used.

The coefficient of variation is:

It is often expressed as a percentage by multiplying by 100. It compares relative variability, but interpretation is problematic when the mean is zero, close to zero or not meaningful, and it is not suitable for variables measured on an arbitrary scale.

Worked example: delivery times

Five delivery times, in days, are 2, 3, 3, 4 and 8.

The mean is (2 + 3 + 3 + 4 + 8) ÷ 5 = 20 ÷ 5 = 4 days. The median is 3 days. The range is 8 − 2 = 6 days. The large value of 8 raises the mean above the median, suggesting right-skewness. If a manager reports only the mean, the typical customer experience may appear slower than the median indicates. Both measures should therefore be reported with the distribution or an explanation of the extreme observation.

6. Exploratory Data Analysis Workflow

A practical EDA workflow is iterative rather than strictly linear.

Step 1: Clarify the decision question

Define the business decision, population, time period and unit of analysis. “Which customers are valuable?” requires a definition of value, such as revenue, margin or lifetime contribution.

Step 2: Audit the data

Check variable names, data types, units, valid ranges, duplicate records, missing values and inconsistent labels. Dates should be genuine dates rather than text strings. Monetary values should use a consistent currency and time basis.

Step 3: Inspect individual variables

Use frequency tables for categorical variables and histograms, box plots or numerical summaries for numerical variables. Look for impossible values, skewness, gaps, multiple peaks and suspicious concentrations.

Step 4: Examine pairs and groups

Use cross-tabulations for categorical pairs, grouped summaries for category–numerical comparisons, and scatterplots for numerical pairs. Compare appropriate rates rather than raw counts when group sizes differ.

Step 5: Investigate time and sequence

Look for trend, seasonality, sudden changes, missing periods and possible changes in data collection. A rise in complaints may reflect a new reporting system rather than a deterioration in service.

Step 6: Validate findings

Recalculate important results, inspect underlying records, compare with an independent source and ask whether alternative explanations exist. Document decisions about exclusions, transformations and definitions.

Missing data

Missingness may be accidental, systematic or informative. For example, income may be missing more often for customers who consider the question sensitive. Reporting the proportion missing and its distribution across groups is essential. Automatically replacing all missing values with zero can create a false measurement.

Outliers and anomalies

An outlier may be a data-entry error, a legitimate rare event, a different population or evidence of an important business process. Analysts should investigate before deleting it. Record the reason for any correction or exclusion and, where practical, conduct a sensitivity analysis with and without the observation.

Transformations

A right-skewed monetary variable may be displayed using a logarithmic axis or transformed variable. If a logarithm is used, the analyst must explain that differences on the transformed scale are not ordinary currency differences. Labels and interpretation must remain clear.

7. Relationships, Correlation and Causation

Covariance describes whether two variables tend to vary together, but its magnitude depends on measurement units. Correlation standardises this relationship. Pearson’s sample correlation is:

Here xᵢ and yᵢ are paired observations, x̄ and ȳ are sample means, and r lies between −1 and +1 when both variables have non-zero variation. The statistic measures linear association and can be strongly affected by outliers, restricted ranges and non-linear patterns. A correlation near zero does not rule out a non-linear relationship.

Association is not causation. Advertising expenditure and sales may be positively associated because advertising increases sales, because growing firms can afford more advertising, or because both increase during a seasonal period. Confounding variables, reverse causality and selection bias must be considered. A scatterplot, correlation and regression line can support exploration, but causal claims require an appropriate research design or strong identifying assumptions.

8. Dashboards and Communication

A dashboard is a coordinated visual interface designed to monitor or explore a defined set of measures. It should support a decision, not merely display every available metric.

A useful dashboard usually includes:

• a clearly stated purpose and audience;
• a small number of priority measures;
• definitions, units and time periods;
• filters that do not create misleading subgroup comparisons;
• consistent scales and colour meanings;
• a visible indication of data freshness;
• links or notes describing sources and calculations;
• alerts that distinguish thresholds from forecasts.

A key performance indicator should have a numerator, denominator where relevant, time window, target and owner. “Conversion rate” is incomplete unless the analyst defines conversions and the population exposed to the opportunity.

Dashboard layout should follow an information hierarchy: context first, priority indicators next, diagnostic detail afterward. Red–green colour schemes may be inaccessible to some users and may imply good/bad judgements where no such judgement is justified. Colour should be supplemented with labels, symbols or position.

A concise analytical communication often follows this structure:

1. Finding: state the important pattern.
2. Evidence: give the relevant comparison, rate or visual.
3. Qualification: state uncertainty, data limitations or alternative explanations.
4. Action: identify a reasonable next investigation or decision.

9. Misleading Visualizations

A technically accurate dataset can still be communicated deceptively. Common problems include:

• truncating a bar-chart axis so small differences appear large;
• using a three-dimensional chart that distorts length and area;
• comparing totals for groups of very different sizes;
• using dual axes that imply a relationship without justification;
• selecting a time window that hides an inconvenient trend;
• omitting zero values or relevant categories;
• using unequal intervals or inconsistent bin widths;
• displaying percentages without stating the denominator;
• using area or volume to represent values, which exaggerates differences;
• suppressing uncertainty, missingness or revisions;
• presenting a correlation as evidence of causation.

Axis truncation is not always wrong. It can be appropriate for a line chart showing small changes, provided the axis is clearly labelled and the presentation does not pretend that the visual length represents a full zero-based magnitude. Bar charts, however, generally require a common zero baseline because bar length is the encoded quantity.

10. Detailed Applied Case: Online Retail Returns

Background

The following is an illustrative case. An online retailer wants to reduce product returns without reducing customer satisfaction. The analytics team receives 48,000 order-line records from one year, including customer segment, product category, delivery days, discount, selling price, return status and satisfaction score.

Problem

Senior management initially asks: “Which product category has the highest return problem?” A simple count could be misleading because categories have different sales volumes.

Analysis

The team first verifies that an order line, rather than a customer, is the observational unit. It removes exact duplicate records but retains legitimate multiple purchases. It checks missingness and finds that satisfaction is absent for 18% of non-returned orders and 7% of returned orders. This difference is documented because complete-case satisfaction comparisons could be biased.

A frequency table shows that electronics has 2,400 returns, more than any other category. However, electronics also has 16,000 orders. The return rate is:

Electronics has 2,400 ÷ 16,000 × 100 = 15%. Home accessories has 1,800 returns among 6,000 orders, giving 30%. Thus, electronics has the largest count, while home accessories has the highest rate.

The team creates a bar chart of return rates with category order fixed from highest to lowest, a box plot of delivery days by return status, and a monthly line chart. The box plot suggests that returned orders have longer delivery times, but several extreme delays occur during a warehouse transition. The monthly chart shows a return-rate increase after a packaging change.

A cross-tabulation by customer segment shows that new customers return 22% of eligible orders and existing customers return 12%. Because the groups differ in product mix, the team avoids claiming that customer age alone causes returns. It examines category and segment jointly and reviews packaging records.

Approach and solution

The retailer pilots stronger packaging for home accessories in two comparable fulfilment centres while maintaining the existing packaging in two others for one quarter. The pilot is not treated as definitive proof unless allocation and other operational changes are assessed. During the pilot, analysts monitor return rate, damage-related returns, customer satisfaction, cost per shipment and delivery time.

The dashboard displays both counts and rates, sample sizes, missingness notes and a data-refresh timestamp. Categories with fewer than a specified minimum number of orders are flagged for cautious interpretation rather than ranked conclusively.

Outcome

The exploratory analysis changes the decision from “reduce electronics returns” to “investigate the high rate of home-accessory returns, especially after the packaging change.” It also identifies delivery delays as a possible contributing factor and prevents the team from treating incomplete satisfaction data as representative.

Learning points

1. Counts and rates answer different questions.
2. The denominator must be visible.
3. Outliers may identify process failures rather than nuisance values.
4. EDA generates hypotheses; it does not automatically prove causes.
5. A dashboard is more useful when definitions, sample sizes and limitations are visible.

11. Ethics, Governance and Reproducibility

Analysts have a duty to avoid selective presentation and unsupported claims. Data visualizations should protect confidentiality, especially when small groups could be re-identified. Aggregation, suppression rules and access controls may be necessary. Personal data should be collected and used for legitimate, transparent purposes, consistent with applicable privacy law and organisational policy.

Governance requires data lineage: where a measure came from, how it was calculated, when it was refreshed and who approved its definition. Version-controlled analysis scripts or documented spreadsheet procedures improve reproducibility. An analyst should preserve the original data, record transformations and distinguish observed facts from interpretations.

AI-assisted visualization tools can accelerate chart selection, narrative generation and anomaly detection. They can also propagate incorrect aggregations, invent explanations or obscure the source of a metric. Human review remains necessary. Sensitive data should not be uploaded to an unapproved service, and generated narratives must be checked against the underlying records.

12. Best-Practice Checklist

Before publishing a chart or dashboard, ask:

• Is the business question explicit?
• Is the observational unit clear?
• Are the denominator, units and time period shown?
• Is the chart type appropriate for the variable types?
• Are categories, axes, legends and colours unambiguous?
• Have missing values, duplicates and outliers been investigated?
• Are comparisons fair in scale and group size?
• Does the display separate association from causation?
• Is uncertainty or data quality information relevant and visible?
• Can another analyst reproduce the result?
• Could any group be harmed, exposed or unfairly characterised?

Chapter Summary

Data visualization and exploratory data analysis are foundational business-analytics practices. Data types and measurement levels guide the selection of tables and charts. Frequency tables and cross-tabulations organise categorical evidence, while histograms, box plots and scatterplots reveal numerical distributions and relationships. Means, medians, standard deviations and interquartile ranges summarise data, but each has assumptions and limitations.

EDA involves clarifying the decision, auditing data quality, examining individual variables, exploring relationships and time patterns, investigating anomalies, and validating findings. Counts must be distinguished from rates, and correlation must not be presented as causation. Dashboards should focus on decisions, define metrics and display denominators, time periods, freshness and limitations. Ethical practice requires transparency, privacy protection, reproducibility and careful review of AI-assisted outputs.

Key Terms

Anomaly: An observation or pattern that differs markedly from expected behaviour.

Association: A tendency for two variables to vary together.

Bar chart: A chart using separated bars to compare categories.

Box plot: A display based on quartiles, median and potential extreme values.

Categorical variable: A variable consisting of labels or groups.

Correlation: A standardised measure of linear association between two numerical variables.

Cross-tabulation: A table showing counts or percentages for combinations of categorical variables.

Dashboard: A coordinated visual display designed to support monitoring or decisions.

Denominator: The reference total used to calculate a rate or percentage.

Exploratory data analysis: Systematic examination of data using summaries, visualizations and quality checks.

Frequency: The number of observations in a category or interval.

Histogram: A chart showing the distribution of numerical data across adjacent intervals.

Interquartile range: The difference between the third and first quartiles.

Missing data: Values that are unavailable, unrecorded or not applicable.

Outlier: An observation unusually distant from other observations or from a modelled pattern.

Scatterplot: A graph of paired numerical observations.

Standard deviation: A measure of dispersion around the mean.

Time series: Observations ordered by time.

Visualization: A graphical representation that maps data to visual properties.

Review Questions

Short Answer Questions

1. Why is the observational unit important when counting records?
2. Distinguish a histogram from a bar chart.
3. What information does the interquartile range provide?
4. Why should a return count often be accompanied by a return rate?
5. State one reason why correlation does not establish causation.

Long Answer Questions

1. Explain how an analyst should conduct an EDA workflow from defining a business question to validating findings.
2. Compare the mean, median, standard deviation and interquartile range, including situations in which each is useful.
3. Discuss how dashboard design can improve or weaken managerial decision-making.
4. Explain at least five ways a visualization can mislead an audience and how each problem can be reduced.
5. Discuss the ethical and governance responsibilities involved in publishing business visualizations.

Analytical / Application Questions

1. A branch records 900 sales and 90 returns; another records 200 sales and 30 returns. Which branch has more returns, and which has the higher return rate? Explain why both comparisons matter.
2. A scatterplot shows a strong positive relationship between advertising expenditure and sales during a holiday season. Identify at least two alternative explanations and propose additional analysis.
3. Design a dashboard for monitoring delivery performance. Specify four measures, their denominators or definitions, suitable visualizations, and two data-quality warnings that should be displayed.

Multiple Choice Questions with Answers

Question 1

Which display is generally most appropriate for comparing sales across separate product categories?

A. Bar chart

B. Histogram

C. Scatterplot

D. Line chart connecting unordered categories

Correct Answer: A. Bar chart

Question 2

A retailer has 24 returns among 60 new-customer orders. What is the return rate?

A. 10%

B. 24%

C. 40%

D. 60%

Correct Answer: C. 40%

Question 3

What does the interquartile range equal?

A. Q₁ + Q₃

B. Q₃ − Q₁

C. Maximum − minimum

D. Mean ÷ standard deviation

Correct Answer: B. Q₃ − Q₁

Question 4

Which statement about a potential outlier is most appropriate?

A. It must always be deleted

B. It proves the data are invalid

C. It should be investigated before a decision is made

D. It should automatically be replaced by the mean

Correct Answer: C. It should be investigated before a decision is made

Question 5

A histogram differs from a bar chart because a histogram generally:

A. Uses only categorical variables

B. Has adjacent bars for numerical intervals

C. Cannot show distributions

D. Must use a secondary axis

Correct Answer: B. Has adjacent bars for numerical intervals

Question 6

If a cross-tabulation reports the percentage of returns that came from new customers, the denominator is:

A. All new-customer orders

B. All orders

C. All returns

D. All existing-customer orders

Correct Answer: C. All returns

Question 7

Pearson’s correlation is primarily a measure of:

A. Linear association

B. Causal impact

C. Data completeness

D. The median difference

Correct Answer: A. Linear association

Question 8

Which dashboard feature most directly improves metric interpretation?

A. Removing all labels to reduce clutter

B. Showing definitions, units and time periods

C. Using three-dimensional bars

D. Hiding the data-refresh date

Correct Answer: B. Showing definitions, units and time periods

Question 9

Five delivery times are 2, 3, 3, 4 and 8 days. What is their median?

A. 2 days

B. 3 days

C. 4 days

D. 8 days

Correct Answer: B. 3 days

Question 10

Which practice is most likely to produce a misleading comparison?

A. Showing both counts and rates

B. Reporting the denominator

C. Comparing raw totals for groups of very different sizes

D. Labelling the time period

Correct Answer: C. Comparing raw totals for groups of very different sizes