Data Mining Models

Data mining is the process of analyzing large amounts of raw data to discover useful information, patterns, and relationships that support better decisions. It is applied across business intelligence, market analysis, political forecasting, web ranking prediction, and weather forecasting — anywhere large volumes of recorded data exist and someone needs to act on what is inside them.

But "analyzing data" is not one activity. A retailer asking which products sell together and a bank asking will this applicant repay are doing fundamentally different things, and they need different kinds of models. A data mining model is the method chosen to answer a specific question from data, and choosing the right one starts with knowing which of two families the question belongs to.

What Is a Data Mining Model?

A data mining model is a method applied to data in order to answer a specific question or solve a specific problem. It takes raw records and turns them into something an organization can use.

Two examples make the idea concrete.

A regression model studies past data and produces a mathematical formula that estimates future values. Financial analysts use regression models to project stock prices and market trends from historical movements.

An association rule model finds items that frequently occur together. If analysis of a stationery retailer's receipts shows that books, pens, and markers are commonly purchased in the same visit, the store manager can place those products near one another to increase sales.

Notice the difference between them, because it is the organizing idea of this whole lesson. The regression model produces an estimate of something not yet known. The association rule model describes a pattern already present in the data. That distinction splits data mining models into two categories.

The Two Categories of Models

The row about labeled outcomes matters most in practice. A predictive model can only be built if you already have historical records where the answer is known — past loan applications with the repayment result recorded, past patients with the diagnosis recorded. If no such labels exist, a predictive model cannot be trained at all, and a descriptive model is the only option available.

Predictive Data Mining Models

Predictive models use historical data to estimate future or unknown outcomes. They identify patterns in past records and apply those patterns to new cases.

Typical applications:

• Healthcare — estimating a patient's risk of conditions such as diabetes or heart failure from clinical history.
• Insurance — estimating accident risk for a policyholder, which feeds into pricing.
• Banking — estimating whether a loan applicant is likely to repay.

Read those carefully and one point stands out: the model produces an estimate of probability, not a certainty. A model that flags a patient as high-risk is saying that people with similar recorded characteristics have historically developed the condition more often. It is not a diagnosis, and treating a predictive output as fact is the most common misuse of these models.

Predictive models come in several forms.

1. Classification

Classification places a record into one of a set of predefined categories, based on patterns learned from records whose category is already known.

The defining feature is that the categories exist before the model is built. Someone decides in advance that transactions are either "fraudulent" or "legitimate", and the model's job is only to decide which label a new transaction gets.

Examples:

• Detecting fraudulent transactions in banking.
• Predicting whether a loan application will be approved or rejected.
• Deciding whether an email is spam.

2. Regression

Regression predicts a continuous numerical value rather than a category. This is the difference that beginners most often blur: classification answers which group, regression answers how much.

• "Will this customer churn?" → classification (yes/no)
• "How many months until this customer churns?" → regression (a number)

Regression studies the relationship between:

• the dependent variable — the output being predicted, and
• one or more independent variables — the inputs used to predict it.

Simple Linear Regression

Simple linear regression uses one independent variable and finds the best straight line describing its relationship with the output.

price = b₀ + b₁ × size

Here b₀ is the intercept — the predicted price when size is zero — and b₁ is the slope, the amount the predicted price changes for each additional unit of size. "Best" line means the one that minimises the total squared distance between the actual data points and the line.

Example: predicting house price from floor area alone.

Multiple Linear Regression

Multiple linear regression uses two or more independent variables:

price = b₀ + b₁ × size + b₂ × rooms + b₃ × age

Instead of fitting a line through a two-dimensional scatter of points, the model fits a plane or hyperplane through a multi-dimensional space — one dimension per input attribute.

Example: predicting house price using floor area, location, and number of rooms.

Note on terminology: multi-linear regression, multiple linear regression, and multivariable regression all refer to this same method — linear regression with more than one input. They are not three different techniques.

Adding variables almost always improves the fit on the training data, which is why fit alone is a poor reason to add them. A variable earns its place only if it improves predictions on data the model has not seen.

Why "Linear"

The word linear refers to how the coefficients combine, not to the shape of the data. A model can capture a curve — by including a squared term, for example — and still be a linear regression, because the coefficients are still added together rather than multiplied by one another.

3. Time Series Analysis

Time series analysis works with data collected at successive points in time, where the order of the observations carries information.

This is what separates it from ordinary regression. In a table of house prices, shuffling the rows changes nothing. In a table of daily sales, shuffling the rows destroys the trend, the weekly cycle, and the seasonal pattern — all of which are exactly what the model is trying to capture.

A time series is usually decomposed into components:

• Trend — long-term upward or downward movement.
• Seasonality — patterns that repeat on a fixed cycle, such as higher retail sales each December.
• Cyclical variation — longer swings not tied to a fixed calendar period.
• Irregular variation — the remaining random fluctuation.

Examples: stock price forecasting, weather prediction, and sales forecasting.

Time series work also requires a different validation approach. Data cannot be split randomly into training and test sets, because that would let the model learn from future observations to predict the past. The split must respect chronological order: train on earlier periods, test on later ones.

Prediction and Forecasting

"Prediction" is often listed as if it were a separate technique, but it is better understood as the purpose that classification, regression, and time series analysis all serve. Estimating an unknown value is prediction; the technique used depends on what kind of value it is.

• Unknown category → classification
• Unknown number → regression
• Unknown future value in a time-ordered series → time series analysis

This framing also clears up a common mislabelling. Credit card fraud detection is frequently given as an example of "prediction", but it is not forecasting a future value at all — it is deciding whether an individual transaction belongs to the fraudulent class, which is classification. When the model instead learns only what normal behaviour looks like and flags whatever deviates from it, without any labeled fraud examples to learn from, the task is anomaly detection, which is descriptive rather than predictive.

Descriptive Data Mining Models

Descriptive models focus on understanding what is already present in the data. They do not forecast future outcomes; they summarize, group, and expose structure so that people can understand what has happened.

Descriptive analytics supports:

• Data summarization
• Pattern discovery
• Data monitoring and reporting

Because descriptive models need no labeled outcomes, they are usually the first thing applied to an unfamiliar dataset — and often the only thing possible when no outcome has been recorded.

1. Clustering

Clustering groups similar data objects together, so that objects within a cluster are more similar to each other than to objects in other clusters.

The algorithm finds the groups; it does not name them or decide which one matters. Interpreting the clusters is human work.

Example: customer segmentation in marketing, where customers are grouped by purchasing behaviour so that campaigns can be targeted at the segment most likely to respond.

2. Association Rules

Association rule mining finds items that frequently occur together in large datasets.

Example: supermarket transaction analysis may reveal that customers who buy milk often buy cereal in the same visit. Retailers use this to design store layouts, bundle offers, and promotions.

A rule is written as:

{milk} → {cereal}

and is judged by three measures, all of which must be read together:

• Support — how often the combination appears in all transactions. Low support means the rule, however striking, applies too rarely to be worth acting on.
• Confidence — of the transactions containing milk, what proportion also contain cereal.
• Lift — how much more likely cereal is when milk is present, compared with its overall frequency. Lift above 1 indicates a genuine association; lift near 1 means the two items merely both happen to be popular.

Confidence on its own is misleading. If 80% of all baskets contain bread, then a rule "{anything} → {bread}" will show 80% confidence while revealing nothing at all. Lift is what catches that.

3. Sequence Discovery

Sequence discovery identifies patterns that occur in a specific order over time. Where association rules ask which items appear together, sequence discovery asks which items appear after which.

Example, from customer purchase behaviour:

Buy phone → Buy phone case → Buy screen protector

The order is the finding. Knowing that screen protectors tend to follow phone purchases tells a retailer when to make the recommendation, not just what to recommend — and that timing is information a plain association rule cannot provide.

4. Summarization

Summarization provides a compact overview of a large dataset, converting complex data into an easily understood form such as reports, charts, or dashboards.

It is the least algorithmically sophisticated technique here and often the most used, because a well-chosen summary frequently answers the actual business question without any modelling at all. Summarization is also what makes the output of every other model communicable to people who will not read the model itself.

Choosing Between the Two

The decision is usually settled by two questions:

1. Is there a specific outcome you want to estimate? If yes, and if historical records of that outcome exist, use a predictive model.
2. If not — or if no labeled outcomes exist — what structure is in the data? That is a descriptive question.

In real projects the two are used in sequence more often than as alternatives. Clustering customers (descriptive) frequently comes first, followed by a separate churn model trained for each segment (predictive). The descriptive step reveals that the groups behave differently; the predictive step exploits that.

Common Mistakes

• Treating a predictive output as a certainty. These models produce estimates from historical patterns, and they are wrong on individual cases routinely.
• Confusing classification with regression. Categories require classification; numeric quantities require regression. Forcing a number into arbitrary buckets to use a classifier discards information.
• Reading confidence without lift. A high-confidence rule about a very common item is usually meaningless.
• Shuffling time series data before splitting it. Random train/test splits let the model see the future, producing excellent test scores and useless real-world performance.
• Expecting clusters to be self-explanatory. Clustering produces groups, not labels or recommendations.

Related Concepts

The techniques named here are covered individually elsewhere in this series: classification and clustering in the lesson on data mining techniques, association rule mining in the Apriori algorithm lesson, and the overall workflow that surrounds model building in the data mining implementation process lesson.