DEV Community

Aaroophan Varatharajan
Aaroophan Varatharajan

Posted on Originally published at aaroophan.medium.com on

Not Every Pattern Deserves a Decision: Rethinking Business Intelligence #5

A predictive model reports impressive accuracy. Management trusts it. Later, the company discovers that it consistently missed the customers most likely to leave. What went wrong?

Finding a statistical pattern is not equivalent to understanding a business problem. A model becomes valuable only when its assumptions, errors, and outputs are appropriate for the decisions it supports.
Finding a statistical pattern is not equivalent to understanding a business problem. A model becomes valuable only when its assumptions, errors, and outputs are appropriate for the decisions it supports.

A supermarket notices that customers who buy potatoes and onions often buy hamburger meat too. Market-basket analysis can uncover that sort of association, and the next move seems obvious: put the products together, adjust a price, or build a promotion around the combination.

It’s a reasonable idea. The relationship came from transaction data, not someone’s intuition.

We have a pattern.

We have a possible action.

But we have quietly ✨skipped✨ a question: does finding a pattern tell us which decision will improve the business?

An association describes what tends to occur together. It does not, by itself, establish why the purchases occur together, whether changing shelf placement will help, or whether a discount will earn back its cost. The algorithm may be doing exactly what we asked. The decision is asking it to do something else.


Discovery Is Not the Same Thing as an Answer

This is the appeal of data mining : instead of specifying every relationship beforehand, we let algorithms search existing information for patterns that may not be obvious. Knowledge Discovery in Databases (KDD) is the broader tradition behind this work; data mining is closely associated with discovering useful patterns within it.

There is a useful distinction between two ways of approaching analysis. In a hypothesis-driven analytical exercise, we begin with a proposition and look for evidence to test it. In exploratory data mining, we may begin with a business interest but without a specific relationship to confirm. We look for what the records can reveal.

Neither is automatically superior, and business analytics overlaps heavily with data mining. The distinction is one of emphasis: data mining leans toward pattern-discovery methods; business analytics asks what those methods mean for a business problem.

The danger is not discovering an unexpected pattern. It is treating interesting as a synonym for actionable.

A retailer can discover co-purchases, an insurer can search historical claims for suspicious patterns, and a telecommunications company can identify customers whose behaviour resembles that of people who previously left. Those are different questions. They require different outputs, different evidence, and different ways of deciding whether the result is useful.


The Business Question Has to Come Before the Model

If the discovery itself cannot justify the decision, then the process cannot sensibly begin with picking the cleverest algorithm.

That is why the Cross-Industry Standard Process for Data Mining (CRISP-DM) begins with Business Understanding, not modeling. The six phases are Business Understanding, Data Understanding, Data Preparation, Modeling, Evaluation, and Deployment. They are not a rigid waterfall; results can send us back to earlier phases.

The Business Question Has to Come Before the Model

The first phase asks what the organization is trying to achieve and how success will be recognized. The next two ask what the available data actually represents, whether it is reliable, and how it must be selected, cleaned, transformed, integrated, and formatted before analysis.

That preparation can feel like the least glamorous part of the work. It is also where the original business question can start to fall apart: a required variable may be missing, two sources may disagree, or the records may not support the proposed prediction. Training a more sophisticated model does not repair a misunderstood input.

Only then does modeling make sense. And even after a model produces a result, evaluation must ask whether it meets the business objective, not merely whether the software ran successfully. IBM’s CRISP-DM guidance describes the same iterative dependency.


Not All Patterns Are Trying to Predict Something

The next trap is choosing techniques because their names sound impressive rather than because their outputs match the question.

Descriptive data mining looks for structure in what has been observed. Association or affinity analysis finds relationships among events or items, as in market-basket analysis; Apriori and Generalized Rule Induction are examples of techniques used to find such rules. Clustering groups similar records without requiring predefined customer labels. A retailer might study buying patterns; a company might use clusters to understand customer segments for marketing.

Harrah’s Entertainment illustrates the business side of that distinction: an enterprise-wide warehouse and customer relationship management analytics helped the company study customers and support its loyalty and service strategy. The point was not to collect clusters for decoration. It was to make customer information usable.

Predictive data mining instead uses historical examples to estimate an outcome for another case. The approaches can be grouped into three families: statistical techniques such as regression; connectionist methods , including artificial neural networks; and symbolic or rule-induction methods , including decision trees.

Not All Patterns Are Trying to Predict Something

The distinction matters because a group of similar customers is not the same output as a prediction that a particular customer will leave. A sales forecast is not a market-basket rule. The goal, available input variables, and required output variable constrain the method. IBM’s overview of data mining makes the same descriptive-versus-predictive distinction.


A Decision Tree Can Explain Its Rules, Not Their Consequences

Consider a targeting model built from historical customer purchases. A decision tree partitions the records through successive tests on predictor variables. In the targeting example, Age in Years provides an important split, and Income is used to divide a subset further.

A Decision Tree Can Explain Its Rules, Not Their Consequences

Algorithms such as CART, CHAID, and C5.0 can induce rules from data. Each internal node tests a variable, each branch represents a result of that test, and the path toward a leaf yields a classification or prediction. This makes a tree comparatively inspectable: its decisions can often be translated into ordinary if/else conditions.

But inspectable is not synonymous with correct. The split points are learned from historical examples. If the examples do not represent the decision the organization now faces, readable rules can still mislead. A neural network may capture relationships that a simple tree cannot; a regression model may suit a different target. Complexity and interpretability are trade-offs to evaluate, not trophies to collect.


The Accuracy Number Is Hiding Two Different Mistakes

Here is where the original confidence in a discovered pattern becomes genuinely uncomfortable.

In a heart-disease classification example, the model correctly classifies 214 out of 303 cases: approximately 70.6% accuracy. That is a legitimate summary of the predictions. Yet the underlying confusion matrix tells us considerably more:

  • True Negatives (TN): 118 — Actual no disease, predicted no disease.
  • False Positives (FP): 46 — Actual no disease, predicted disease.
  • False Negatives (FN): 43 — Disease present, predicted no disease.
  • True Positives (TP): 96 — Disease present, predicted disease.

Accuracy = (118 + 96) / 303 ≈ 70.6%

The same model makes 46 false-positive predictions , flagging disease when none is present, and 43 false-negative predictions , missing disease that is present. Accuracy counts both as one mistake each. Their practical consequences need not be equal.

For a medical diagnostic decision, missing a patient who actually has a disease may be more consequential than referring a healthy patient for further examination. The model’s recall for disease is 96 / (96 + 43) ≈ 69.1%: it identifies roughly seven out of ten disease-positive cases in this example. The overall accuracy does not tell us that directly.

This is not an argument that false positives never matter. They can also cause harm, cost, and unnecessary follow-up. The point is that the cost of each error depends on the decision being supported. Changing a classification threshold may trade one error type against another, so evaluation needs to consider that trade-off deliberately. Scikit-learn’s confusion-matrix documentation describes the four outcomes precisely.

And now the hidden constraint is visible. We cannot select a model using a single attractive score unless that score reflects what the business, or in this case, the clinical decision, actually needs.


A Correct Prediction Still Needs Somewhere to Go

Even a model that performs well on an appropriate test dataset can become an expensive ornament if nobody knows how to use its output.

This is why evaluation and interpretation are followed by deployment. Interpretation helps users understand what a model’s recommendation means. Deployment makes that result available in the workflow where The Origami Software Engineer can respond to it. CRISP-DM allows a return to data preparation, modeling, or even business understanding when evaluation reveals a mismatch; it does not demand that a disappointing model be shipped anyway.

The implementation pitfalls are painfully ordinary: exaggerated expectations, sophisticated tools assigned to the wrong users, results delivered without an action process, training that ends after launch, and quick wins pursued without a sustainable information foundation. Weak data governance and an inability to demonstrate business value make the gap wider.

Real-time decision support adds another constraint. Analytics needs to be interactive enough to respond to changing situations and sufficiently invisible in routine use that decision-makers need not operate a separate analytical laboratory. That means addressing large datasets and high-dimensional models, automated discovery, understandable outputs, metadata, and integration with existing environments and organizational levels.

The system should not simply announce, “Here is a pattern.” It should put the relevant insight where an appropriate decision can be made. IBM’s deployment guidance likewise distinguishes putting a scoring model into operation from using discovered insights to change how an organization works.


The Decision Was Never Inside the Pattern

Return to the supermarket. The co-purchase relationship was useful evidence. It could motivate a question about shelf arrangement or promotions. It could even justify further analysis or a controlled experiment. What it could not do, on its own, was guarantee that the proposed change would be profitable.

The same boundary appears with customer clusters, purchase predictions, and diagnostic classifications. A pattern has a particular meaning; a model has particular inputs and errors; a decision has consequences. Those are related, but they are not interchangeable.

Once we accept that, the apparently lengthy route through business understanding, preparation, modeling, evaluation, and deployment stops looking like ceremony. Every stage protects the decision from a different kind of misplaced trust.

Data mining can tell us what it found. Business Intelligence still has to establish whether that finding deserves to change what we do.

That’s not failure.
That’s evolution.


The “I liked this” Starter Pack:

Don’t let your fingers get lazy now.

  • Like : It tells me this was worth writing.
  • A Comment: Tell me your thoughts, your favorite snack, or a better title for this blog.
  • Boost it: Especially with that one developer who definitely needs this.

Thanks for being here. It genuinely helps more than you know!

— Aaroophan Varatharajan

Find me elsewhere:

Top comments (0)