Descriptive and Predictive Data Mining

Data mining techniques discover patterns and useful information in large datasets. They divide into two families, distinguished by the question they answer.

Descriptive data mining analyzes past data to understand what has already happened, identifying patterns, relationships, and trends in historical records.

Predictive data mining uses past and present data to estimate future outcomes, helping organizations decide by forecasting what is likely to happen.

Put simply:

• Descriptive mining explains patterns and relationships in existing data.
• Predictive mining uses those patterns to make statements about the future.

The order matters. Descriptive work almost always comes first, because you cannot sensibly forecast something you do not yet understand.

What Is Descriptive Data Mining?

Descriptive data mining summarizes and interprets historical data, helping to understand the characteristics and patterns within a dataset.

Common techniques include:

• Correlation analysis — measuring how strongly two variables move together
• Cross-tabulation — counting how records are distributed across combinations of categories
• Frequency analysis — counting how often each value or event occurs

These identify regularities in the data, converting raw records into meaningful information that can be used for reporting and monitoring.

Descriptive analytics answers the question:

"What happened in the past?"

By understanding past events, organizations can learn from their data and improve their future strategies.

Why It Comes First

The practical reason descriptive mining precedes predictive mining is that it does not require labeled outcomes.

A predictive model needs historical records where the answer is already known — past customers with the churn result recorded, past applicants with the repayment result recorded. On an unfamiliar dataset that requirement is often not met at all, and even when it is, nobody yet knows which outcome is worth predicting.

Descriptive techniques have no such requirement. They can be applied to any dataset immediately, and what they reveal usually determines which predictive question is worth asking.

Examples in Learning Analytics

Education provides a clear set of descriptive examples:

• Comparing pre-test and post-test results of students
• Tracking course enrollment numbers
• Collecting and analyzing course survey results
• Monitoring how often learning resources are accessed
• Counting the number of discussion posts made by learners

Each of these summarizes something that has already occurred, and together they help educators understand student behaviour and improve the learning process.

Notice that none of them forecasts anything. Knowing that a resource was accessed 40 times is a description. Estimating which students will fail the next assessment is a prediction — and it depends on descriptive measures like these as its inputs.

Descriptive Techniques in Data Mining

Beyond the summary statistics above, the main algorithmic descriptive techniques are:

• Clustering — grouping similar records so natural segments emerge from the data.
• Association rule mining — finding items or events that frequently occur together.
• Sequence discovery — finding events that frequently occur in a particular order.
• Summarization — condensing a large dataset into reports, charts, or aggregate measures.
• Outlier detection — identifying records that do not fit the general pattern.

What Is Predictive Data Mining?

Predictive data mining forecasts future events or trends from historical data, using algorithms, statistical models, and machine learning to analyze past patterns and estimate likely outcomes.

It supports predictive analytics, which answers:

"What is likely to happen in the future?"

Example

A retail store analyzes its customer purchase history. By studying previous transactions, it can estimate:

• What products customers may buy next
• When customers are likely to shop again
• Which products will see higher demand

This helps the business plan inventory, marketing, and sales campaigns more effectively.

It is worth being precise about what such a model produces. It does not say that a particular customer will return on Thursday. It says that customers whose past behaviour resembles this one have historically returned within a week, so this customer probably will too. That is a likelihood, and treating it as a certainty is the most common way predictive results get misused.

Predictive Techniques in Data Mining

• Classification — assigning a record to one of several predefined categories.
• Regression — estimating a continuous numeric value.
• Time series analysis — forecasting future values in an ordered sequence of observations.

All three are supervised techniques: they require historical records where the outcome is known, so that the algorithm has something to learn from.

Advantages of Predictive Data Mining in Business

• Helps increase business productivity
• Reduces business risk by anticipating problems
• Supports better decision-making for analysts and managers
• Helps companies maintain a competitive advantage

Side-by-Side Comparison

Where They Sit in the Analytics Maturity Model

Analytics work is often described in four levels of increasing difficulty and value. Placing descriptive and predictive mining within it clarifies both what they do and what they do not.

Two points follow from this table.

First, descriptive mining covers more than the first level. Correlation analysis and clustering are frequently what answer the diagnostic question of why something happened, which is why "descriptive" should not be read as "merely reporting".

Second, neither descriptive nor predictive mining reaches the prescriptive level. A model can estimate that a customer will churn; deciding whether to offer a discount, call them, or let them go is a business decision that no algorithm in this lesson makes.

How They Work Together

A typical project uses both, in sequence.

1. Descriptive. Clustering the customer base reveals four distinct segments, one of which has an unusually short average relationship length.
2. Diagnostic. Cross-tabulation shows that this segment was acquired mainly through one marketing channel.
3. Predictive. A classification model, trained on historical churn records, scores each current customer's likelihood of leaving within 90 days.
4. Action. The retention team works the high-scoring customers in the affected segment.

Step 3 was only possible because steps 1 and 2 established what was worth predicting. Running the classification model first — before anyone knew the segments existed — would have produced a churn score with no idea of what to do with it.

Common Mistakes

• Treating descriptive mining as merely reporting. Clustering and association rules discover structure nobody specified; that is a great deal more than counting.
• Jumping to prediction on an unfamiliar dataset. Without descriptive work first, the predicted quantity is usually the wrong one.
• Reporting a prediction as a fact. Predictive models produce likelihoods that are wrong on individual cases routinely.
• Attempting predictive mining without recorded outcomes. If history does not record what actually happened, there is nothing to learn from.
• Expecting either family to say what to do. Both produce findings; the decision remains human.

Related Concepts

This split is developed further in this series' lessons on data mining models and types of data mining. The individual techniques named here — clustering, association rules, classification, and regression — each have their own lesson, and the relationship with predictive analytics as a business discipline is covered in the predictive analytics lesson.