Preprocessing

Definition

Preprocessing is the preparation and transformation of data before it is used by a model or another analysis step. It can include checking formats, handling missing values, scaling measurements, encoding categories and constructing a consistent input representation.

What it means

Raw records rarely arrive in exactly the form a model needs. Dates may use different formats, categories may be written inconsistently, and measurements may use incompatible units. Preprocessing establishes how those records become usable inputs.

Some operations follow a fixed rule, such as converting centimeters to meters. Others learn information from a dataset, such as a mean used for imputation or a vocabulary used for encoding text. That distinction matters when evaluating a model.

Preprocessing is broader than deep learning and is not automatically beneficial. Removing unusual values can erase important cases; simplifying text can remove meaning. Each transformation should have a purpose tied to the data and task, with enough documentation to reproduce the same interpretation later.

How it works

Start by defining an input schema: expected fields, types, units and permitted missing values. Check whether the data meets that contract and decide how exceptions should be handled.

Then apply appropriate transformations. Numerical features may be scaled; categorical values may be encoded; missing values may receive a documented treatment. Different model families have different requirements, so one universal recipe is inappropriate.

For an ordinary held-out evaluation, fit learned transformations using the training subset only, then apply the fitted transformation to validation or test inputs. During cross-validation, fitting belongs inside each training fold. Otherwise information from evaluation data can influence the model indirectly.

Save the fitted preprocessing state with the model and use the same interpretation at prediction time. Handle unseen categories, unexpected units and missing fields explicitly. A pipeline can help keep the steps together, but it does not decide whether their underlying assumptions are sensible.

Three practical examples

A demand model converts all recorded quantities to a common unit and applies a scaler fitted on its training data.
A service-ticket classifier uses a consistent text representation when learning from historical messages and processing new ones.
A sensor-analysis workflow identifies missing readings, retains a missingness indicator where useful and applies a documented imputation method.

Think of it this way

Think of preparing ingredients to a recipe's specified measurements before cooking. Converting units and deciding what to do with a missing ingredient can determine whether the recipe works as intended.

The analogy does not imply that data should always be made uniform or that unusual records are spoiled. In some tasks, differences and rare cases carry the most important information. Preparation must preserve what the task needs rather than merely making the inputs look tidy.

Worked example

Fit once, then apply consistently

Suppose a training feature contains the invented values 10 and 20. A min-max transformation fitted on this training set uses minimum 10 and maximum 20, so it maps a value x to (x − 10) / 10. The training values become 0 and 1.

A later value of 30 becomes 2 under that same transformation. Being outside the original zero-to-one range is possible when a new value exceeds the training range; clipping would be a separate design choice.

If the training and later value were combined before fitting, the maximum would become 30 and the original training value 20 would instead map to 0.5. Evaluation information would have changed the representation used to learn. If the later data were independently refitted, its scale would no longer match what the model expects. The correct treatment depends on the stated evaluation and deployment protocol.

Illustrative case study

Fictional example: Birch Maintenance builds a model using temperature readings from several sites. Historical records store all temperatures in Celsius. A new site's feed provides Fahrenheit values in a field with the same name.

The prediction service successfully parses the numbers and reports no software error, but its outputs shift sharply. Retraining a larger model would not resolve the ambiguous input contract. The team adds explicit unit metadata and checks the conversion before applying the saved training transformations.

It also revisits missing readings that were previously replaced with zero. Zero degrees can be a genuine observation, so the revised representation distinguishes absent measurements from measured values.

The case illustrates that reliable preprocessing depends on meaning, not just numerical compatibility. A model can accept an input array while receiving a fundamentally different description of the world.

Common misconceptions and mistakes

Preprocessing means deleting all outliers. Unusual observations may be valid and important.
Unsupervised transformations cannot leak information. Scaling, imputation and dimensionality reduction can still learn from evaluation data.
Test data should be left unprocessed. It should receive the appropriate transformation fitted without using its held-out information.
A successful conversion proves data quality. Correct types do not establish correct units, labels or measurement meaning.

Business use

Preprocessing often determines whether a model behaves consistently between development and daily use. Treat it as part of the product's data contract, with ownership for changes in source systems and interpretation.

Review how many records each step changes or excludes, and examine whether exclusions disproportionately remove a relevant class of cases. Preserve the versioned transformation settings needed to reproduce predictions. When model performance changes, inspect the input pipeline alongside the model itself. A changed category spelling, unit or missing-value convention can matter more than a change in the prediction algorithm.

Check your understanding

A team fits a missing-value imputer on the entire dataset, then splits the result into training and test sets. Has it kept the test set completely independent of model preparation?

Answer: No. The imputation values were influenced by the test records. Fit the learned transformation within the training portion of the evaluation protocol.

Related terms

Data cleaning
Feature engineering
Data leakage
Pipeline
Training-serving skew
feature scaling
imputation
categorical encoding

Learn next

Read data leakage and feature scaling. List each step in a proposed preprocessing pipeline, marking whether it follows a fixed rule or learns from data, and record where that learned state is fitted and reused.

Sources

scikit-learn developers, preprocessing transformations: https://scikit-learn.org/stable/modules/preprocessing.html. Common pitfalls, inconsistent preprocessing and leakage: https://scikit-learn.org/stable/common_pitfalls.html. Google Machine Learning Glossary, data and feature terminology: https://developers.google.com/machine-learning/glossary. The numerical example and maintenance case are original illustrations.

Previous: Prepositions Next: Prescriptive analytics