Visualization is part of the data-mining workflow, not a decorative final step. Analysts use charts to inspect inputs, detect data-quality problems, explore relationships and clusters, examine model outputs, and communicate conclusions. The right display depends on the question, the structure of the data, and how the apparent pattern will be checked against the underlying records or model.
Where visualization fits in a data-mining workflow
A practical workflow uses visual displays at several points:
As an Amazon Associate I earn from qualifying purchases.
- Before modeling: inspect variable ranges, missing values, unusual records, class imbalance, and possible transformations.
- During exploration: look for comparisons, trends, associations, clusters, gaps, and changes across groups or time.
- After modeling: examine predictions, errors, feature behavior, cluster assignments, and the shape of discovered structures.
- During communication: present the evidence that answers the business or scientific question, with enough context to prevent a misleading interpretation.
A visual pattern is evidence for investigation or validation. It is not, by itself, proof that one variable causes another. Domain knowledge, the mining objective, and the original data must remain part of the interpretation.
Quick wins for a faster PC:
Scan for outdated or missing drivers - takes under a minuteDriver Scan →Clear out junk files and repair common Windows errorsFree Scan →Start with the task and data structure
Choose an encoding that matches both what you need to learn and what the data can represent clearly. The table below gives practical starting points rather than universal rankings.
#1 Best Overall
- Wiley
- Language: english
- Book - storytelling with data: a data visualization guide for business professionals
| Question or task | Typical data shape | Useful display | What it reveals | Main caution |
|---|---|---|---|---|
| How do categories compare? | Categorical variable with one measured value per category | Bar chart | Relative magnitude and rank | Long category lists become difficult to read; inconsistent baselines can exaggerate differences. |
| How does a measure change over an ordered sequence? | Values indexed by time or another meaningful order | Line graph | Direction, turning points, and sustained changes | Connecting unordered observations implies a continuity that may not exist. |
| Are two numerical variables related? | Paired numerical observations | Scatter plot | Association, nonlinearity, groups, and outliers | Overplotting can hide dense regions; association does not establish causation. |
| What is the distribution of one measure? | Many observations of a numerical variable | Histogram | Shape, concentration, spread, and possible multimodality | Bin width can change the apparent story. |
| How do distributions differ between groups? | Numerical measure split by categories | Boxplot | Median, spread, and unusual observations in each group | Its summary can hide multimodality and sample-size differences. |
| How do several variables behave together? | Many numerical variables measured on common records | Parallel coordinates, radial displays, or a self-organizing map | Profiles, separations, and possible clusters | Readability declines as dimensions and records increase; scaling and ordering affect the display. |
| What is the structure of connected entities? | Nodes and relationships | Network visualization | Hubs, communities, paths, and connectivity | Dense networks quickly become visually cluttered. |
| How is information nested? | Parent-child or containment relationships | Hierarchical visualization | Levels, branches, and relative composition | Area, angle, or depth encodings can make precise comparisons difficult. |
| Where do observations occur? | Geographic coordinates or regions | Geographic map | Spatial concentration, gaps, and regional differences | Projection, geographic area, and unequal population sizes can distort comparisons. |
Basic charts for common mining questions
Bar charts for category comparisons
Use bars when the question is “which category is larger, smaller, or ranked differently?” Sort categories when ranking matters, label units, and keep the baseline honest. If the categories are parts of a whole, state whether the values are counts, rates, or percentages; the same bar shape can support very different conclusions.
Line graphs for ordered change
Lines are appropriate when observations have a meaningful order, most often time. Mark missing periods and distinguish observed values from estimates. Multiple series can expose divergent trends, but too many lines turn comparison into tracing. Small multiples or a focused subset may communicate more clearly.
Scatter plots for relationships
A scatter plot is a first choice for two numerical variables. Use color, shape, or faceting sparingly to show a third categorical variable. Inspect whether an apparent relationship is driven by a few points, a hidden subgroup, a nonlinear shape, or unequal density. When many points overlap, transparency, aggregation, or sampling can help, but each changes what the viewer sees.
The Tool Desk
Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →Histograms and boxplots for distributions
Histograms show how observations are distributed across numeric intervals; changing the bin width can reveal or conceal structure. Boxplots provide compact group comparisons, including a median and a measure of spread, but they do not show every feature of the distribution. For an important decision, pair a summary plot with the underlying observations or a more detailed distribution view.
Visualizing multidimensional data
When each record has many variables, no single two-dimensional chart can preserve every relationship. The following methods named in the textbook chapter Data Mining, third edition, are useful when their assumptions and visual load are made explicit.
Parallel coordinates
Each variable becomes a vertical axis, and each record is drawn as a line crossing the axes at its values. This can expose profiles, groups, and variables that separate categories. Normalize variables when units or ranges differ, order axes to place related variables near one another, and use filtering or brushing to avoid an unreadable mass of lines. Crossings are not automatically meaningful; they depend on axis order and scale.
Radial visualization
Radial displays arrange variables around a circle and encode each record or profile across those axes. They can make a multivariable signature visually distinctive, especially for a small number of profiles. However, comparing distances and angles around a circle is harder than comparing aligned positions, so use radial views for pattern discovery rather than precise measurement.
Self-organizing maps
A self-organizing map places high-dimensional observations on a grid so that nearby cells represent similar profiles. Coloring cells by a variable, class, or cluster assignment can reveal broad regions and transitions. Treat the map as a model-based projection: investigate the original records and variables before treating a visible region as a substantive group.
Use coordinated views when one display is insufficient
Linked charts can combine a scatter plot, distribution view, map, or table. Selecting observations in one view and highlighting them in others helps connect an unusual point to its values, location, category, or model error. This interaction is most useful when it answers a defined investigative question; interaction should not replace clear labels or a reproducible analysis.
Visualizations for structured data
Networks
Use a network view when relationships are the object of analysis: for example, links among entities, citations, or transactions. Choose whether node size, color, edge width, or layout represents a measured quantity, and explain that choice. A visually central node may reflect the layout or the selected metric rather than practical importance.
Hierarchies
Tree diagrams, nested rectangles, and related views suit taxonomies, folders, organizational structures, and other containment data. They show branching and composition, while a companion table or bar chart is often better for exact comparisons.
Geographic data
Maps are appropriate when location changes the meaning of the observation. Distinguish counts from rates, state the geographic unit, and consider whether large regions visually dominate because of area rather than population or exposure. A map should be paired with the denominator and time period used to calculate the displayed value.
Best Value
Interaction, validation, and responsible interpretation
Interactive filtering, zooming, brushing, sorting, and tooltips can support visual data mining by letting analysts move from an overview to the records behind a pattern. Save the filters, transformations, and selections used to produce an important view so another analyst can reproduce it.
- Check the data: verify units, missingness, duplicated records, coding changes, and extreme values.
- Check the model: compare visual patterns with predictions, residuals, validation results, or cluster assignments.
- Check alternatives: vary bin widths, scales, category order, and relevant subsets to see whether the pattern persists.
- Check the context: ask whether the period, population, geography, and domain mechanism support the interpretation.
- Check the communication: label uncertainty, denominators, transformations, and whether values are observed, estimated, or predicted.
John W. Tukey is attributed the observation: “The greatest value of a picture is when it forces us to notice what we never expected to see.” An unexpected visual pattern is a reason to investigate, not a license to skip measurement or validation.
Quick Recap
A repeatable selection procedure
- State the decision or question. Write what the viewer must compare, locate, classify, explain, or discover.
- Identify the data structure. Record whether variables are categorical, numerical, ordered, geographic, hierarchical, or relational, and note the number of observations and dimensions.
- Select the simplest suitable encoding. Start with bars, lines, scatter plots, histograms, or boxplots when they answer the question; move to multidimensional or structural views only when necessary.
- Make scales and units explicit. Document transformations, baselines, binning, normalization, denominators, and time windows.
- Inspect and interact. Filter or link views to examine suspected subgroups, outliers, and individual records.
- Validate outside the picture. Recalculate relevant values, test the mining result with the appropriate evaluation procedure, and use domain knowledge to assess plausibility.
- Publish the view that supports the conclusion. Include the context and limitations a reader needs, rather than every exploratory chart created along the way.
Further reading
- Data Mining, third edition, by Jiawei Han, Micheline Kamber, and Jian Pei (Wiley), includes a “Visualization Methods” chapter covering perception, scientific and information visualization, parallel coordinates, radial visualization, self-organizing maps, and visualization systems for data mining.
- Data Mining: Practical Machine Learning Tools and Techniques, third edition (Elsevier), discusses the Weka toolkit and visualization among its task areas.
- Visual Data Mining (Wiley) presents a visual methodology and exercises associated with the VisMiner tool.
- Information Visualization in Data Mining and Knowledge Discovery (Elsevier) collects chapters on visualization concepts, interaction, model visualization, and data-mining applications.
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
