AI Glossary / AI and machine learning / Intermediate

Representation

Definition

A representation is the form in which data is presented to a model: the encoding of a customer, an image or a sentence as numbers the algorithm can work with, and the choice of that form often matters more than the choice of model.

What it means

Models do not see the world; they see numbers. A representation is the translation layer between reality and arithmetic, deciding what aspects of a thing become measurable and what is thrown away.

Good representations make learning easy. A customer encoded as recency, frequency and value is far more learnable than the same customer as a raw string of transactions, because the pattern is already surfaced in the numbers.

Modern systems increasingly learn their own representations: embeddings turn words, products and images into points in a space where closeness means similarity, discovered from data rather than designed by hand.

How it works

Decide what about each item could possibly matter for the task, and expose it as explicit numeric features.

Normalize and encode: scale numbers, turn categories into indicator columns, keep units honest.

Or let a model learn the representation, training embeddings so similar items land near each other in the vector space.

Test the representation with a simple model first; if simple methods do well, the encoding is carrying the signal.

Reuse learned representations across tasks, since a good embedding of customers or products transfers to new questions.

Visualize learned embeddings of key entities, since nonsense neighbors are an early warning of a broken encoding.

Three practical examples

• Representing customers as recency-frequency-monetary scores for churn work instead of feeding raw order logs.
• Product embeddings learned from purchase baskets, so substitute products sit close together and recommendations improve.
• Representing support tickets as embeddings, so clustering reveals issue themes nobody had named.

Think of it this way

Think of maps. The same city can be drawn as a road atlas, a tube diagram or a contour chart, and each makes some questions trivial and others impossible. The territory is one; the representation decides what is easy to see.

Worked example

the two encodings

Suppose a retailer wants to predict which shoppers will buy a new snack line. The figures are illustrative.

Encoding customers as raw text of every receipt, a strong model achieves barely better than chance; the signal is buried.

Encoding the same customers as snack-category spend, visit frequency and coupon use, a simple model scores far higher, because the pattern is pre-digested into the numbers.

The general lesson: hours spent on representation routinely beat weeks spent on algorithms, and this is the least appreciated lever in applied machine learning.

Illustrative case study

A logistics firm spent a quarter tuning algorithms for late-delivery prediction with little progress. A new analyst instead rebuilt the representation: deliveries were encoded with features for time pressure, route familiarity and handover counts rather than raw GPS trails. The simplest model in the original bake-off then beat everything tried before. The firm changed its staffing rule as a result: representation design goes to the person who knows the operation best, not the person with the fanciest toolkit, because the features are where domain knowledge enters the maths.

Common misconceptions and mistakes

• The model is what matters. Given a weak representation, the best model struggles; given a revealing one, simple models shine.
• Learned representations remove the need for thought. Embeddings encode the biases and blind spots of their training data.
• A representation is neutral. Every encoding chooses what to make visible and what to erase, which is a decision with consequences.
• More features mean a richer representation. Irrelevant columns dilute signal and cost; relevance, not volume, makes a representation rich.

Business use

When a project stalls, change the representation before changing the algorithm.

Put domain experts in charge of feature design sessions; that is where their knowledge becomes mathematical.

Ask what any embedding was trained on before trusting its similarities for decisions.

Keep representations documented and versioned, because downstream models silently depend on them.

Audit representations for sensitive attributes and their proxies before using them in decisions about people.

Budget explicitly for representation work in project plans; it is usually the highest-return line item and the first one squeezed when timelines tighten.

People also ask

What is a representation in machine learning?

The numeric encoding of raw data that a model actually sees, whether hand-designed features or learned embeddings.

Why do learned embeddings matter for business?

They place words, products or customers in a space where closeness means similarity, powering search, recommendations and clustering without hand-built rules.

How do you know a representation is good?

A simple model performs well on it, related items group sensibly, and the encoded facts match what domain experts believe matters.

Who should design representations?

A partnership: domain experts nominate what matters, data people encode it safely, and simple-model tests arbitrate whether the encoding carries signal.

Check your understanding

What is a representation, and why does it often matter more than the model?

Answer: A representation is the numerical form data takes when presented to a model, from engineered features to learned embeddings. It matters more than the model because it decides which patterns are visible at all: a revealing encoding lets simple models excel, while a poor one defeats sophisticated algorithms, so representation design is where domain knowledge enters the mathematics.

Related terms

• Feature engineering
• Embedding
• Feature vector
• Encoding (for Features)
• Transfer learning
• Dimensionality
• encoding

Learn next

Be Productive Using AI shows how to turn raw business material into structured inputs AI tools can use well: https://aiforbusinessandlife.com/academy/be-productive-using-ai [C1]

More in AI and machine learning

• Reparameterization trick - The reparameterization trick rewrites a random sample as a deterministic function of model parameters and parameter-independent noise.
• Replay buffer - A replay buffer is a memory of past experiences that a reinforcement learning agent stores and re-practices on, so each hard-won lesson is learned from many...
• Replica - A replica is an identical running copy of a model or service, deployed alongside others so the system can serve more requests at once and survive the...
• Representation collapse - Representation collapse occurs when a learned representation loses distinctions needed to describe its inputs.
• Return - In reinforcement learning, the return is the total reward an agent collects from a point onward, usually with future rewards discounted, and it is the...
• Robot learning - Robot learning is the use of data and experience to help a robot acquire or improve capabilities.

Sources

[S2] Google, Machine Learning Glossary; [S25] Hugging Face, Transformers Glossary. Full URLs are in the source ledger. Verified 11 October 2026.

Previous: Reporting events Next: Representation Bias