By convention in statistics and machine learning texts, the inputs are X and the outcomes y.

What is a sparse matrix?

One where most cells are zero, common in text and basket data, stored in compressed form.

Can the matrix contain text?

Not directly; text must be encoded into numeric features first.

AI Glossary / AI and machine learning / Beginner

Feature matrix

Definition

A feature matrix is the table a model trains on: one row per example, one column per feature, with each cell holding that example's value for that field.

What it means

Before any learning happens, the data must become a table. Each row is one case, a customer, an order, a day; each column is one feature; and the grid of values is the feature matrix, conventionally called X.

The matrix is the meeting point of business and mathematics. Everything the model may know is in it, and everything outside it is invisible, so the matrix is, quite literally, the world as the model is allowed to see it.

Its companion is the target vector, the column of outcomes the model learns to predict, conventionally called y. Training is the search for a rule that maps X to y.

Matrix hygiene decides model quality: mixed units, silent missing values, duplicated rows and misaligned columns each corrupt learning without crashing anything, so inspecting X is a ritual before any training.

For a business, the feature matrix is also the audit surface: reviewing its columns tells you what the model knows, and reviewing its rows tells you whose behavior it will learn from.

How it works

Assemble rows so each represents one prediction case, at the moment the prediction would be made, using only data available then.

Assemble columns so each is a clean, documented feature: typed, encoded and free of leakage.

Align with the target: every row's outcome value must correspond to that same case, at the right later time.

Inspect the matrix before training: shapes, types, missingness, ranges and duplicates, plus a sample of actual rows.

Version the matrix used for each training run, so any model can be traced back to the exact table that produced it.

Store the matrix in a form that preserves types and provenance; a matrix flattened into anonymous numbers loses the lineage that audits and debugging will later demand.

Three practical examples

Churn model: one row per customer per month, columns for usage, tenure, support contact, and the churn flag as target.
Demand model: one row per product per store per week, columns for price, promotion and seasonality.
Fraud model: one row per transaction, columns for amount, merchant, time and behavioral history.

Think of it this way

Think of the register at an exam board. Each student is a row, each subject a column, each cell a mark. Nothing outside the register can be graded, because the register is the whole of what the board sees.

The limit: registers inherit the school's blind spots. If no one recorded attendance, attendance cannot be graded, and the model will never know it mattered.

Worked example

auditing a training table

Suppose a team assembles a churn matrix of 200,000 customer-months and runs its standard pre-training ritual. The figures are illustrative.

The audit finds 3 per cent of rows duplicated by a join error, one column measured in a mix of dollars and cents, and a date field whose missing values were silently filled with 1970.

Fixed before training, none of these would ever have been visible in the final accuracy number, only in worse decisions.

The team keeps the checklist: duplicates, units, missing-fill sentinels, and a sample of rows read by eye.

Illustrative case study

A fictional lender trained on a matrix assembled with information from after the decision date, because the join pulled the customer's file as of today rather than as of application.

Every accuracy number was inflated and every live decision disappointing, for six months, until a reviewer asked when each column was knowable.

The invented lesson: a feature matrix is a claim about the past. Rows assembled with future knowledge are fiction, and fiction trains beautifully.

Common misconceptions and mistakes

The matrix is the raw data. It is a prepared view: rows chosen, columns engineered, and those choices shape the model as much as the algorithm does.
Bigger matrices are always better. More rows and columns help only when they add information; duplicated rows and empty columns add nothing but cost.
Missing values announce themselves. They hide behind zeros, sentinels and defaults; only deliberate audits surface them.

Business use

Review the feature matrix in model governance as the model's declared worldview: its columns are what the model may know.

Ask of every training table: as of what moment is each row assembled, and is any column knowable only later.

Keep the matrix for every model version, because reproducing a model without its exact training table is archaeology, not engineering.

Treat the assembled matrix as the deliverable of data preparation, reviewed and signed off like a financial statement, because the model is downstream of its accuracy.

Check your understanding

Why is the feature matrix described as the world the model is allowed to see?

Answer: Because the model learns only from the matrix's rows and columns, so any information not assembled into that table is invisible to it, whatever the raw data held.

Related terms

Learn next

Be Productive Using AI builds habits for inspecting the data behind any claim, starting with what is actually in the table: https://aiforbusinessandlife.com/academy/be-productive-using-ai. [C1]

Sources

[S12] Aalto University, Dictionary of Machine Learning. Full URLs are in the source ledger. Verified 11 October 2026.

Previous: Feature map Next: Feature Monitoring