What is a one-hot vector?

A row of zeros with a single 1 marking which category an item belongs to.

How does it relate to one-hot encoding?

The encoding is the scheme; the vector is one item's row under that scheme.

Do one-hot vectors show similarity?

No: every category is equally distant from every other, so similarity must be learned separately, as in embeddings.

Why does position order matter?

Position is identity: the third switch means Paris only by convention, and changing the order renames everything.

Can more than one position be 1?

Not in a true one-hot vector; several 1s make it multi-hot, the representation for multi-label membership.

AI Glossary / AI and machine learning / Beginner

One-hot vector

Definition

A one-hot vector is the row of switches that represents one item's category: all zeros except a single 1 in the position of the category the item belongs to.

What it means

Where one-hot encoding is the scheme, the one-hot vector is the result for a single item: Paris becomes 0,1,0,0 in a four-city world, one switch on, the rest off.

The vector's geometry is neat: every category sits at a corner of the space, equally distant from every other, so no category is accidentally closer to another. Paris and Cairo are as apart as Paris and Oslo.

That equal distance is both honesty and limitation: the encoding invents no false relationships, but it also encodes no real ones, so similarity between categories must be learned from data, not read from the vectors.

One-hot vectors are the starting point for richer representations: embeddings are, in effect, learned replacements that place related categories near each other in a denser space.

For a business, the one-hot vector is the atom of categorical data handling: understanding that every category column is one switch, and every item is one row of mostly zeros, demystifies half of feature engineering.

The vector is sparse by design, one mark in a long row, which is why wide vocabularies push practitioners toward dense embeddings that pack meaning into fewer numbers.

How it works

Assign each category a fixed position in the vector, and keep that ordering forever, since position is identity.

Set the item's own category position to 1 and all others to 0, one switch on per field.

Concatenate the vectors of several categorical fields when an item has many, producing one long row of switches.

Store the mapping from positions to categories with the model, since a bare vector is meaningless without its legend.

Move to embeddings when the space gets wide, trading the explicit switches for learned positions that carry similarity.

Three practical examples

A four-city encoding where Paris is 1,0,0,0 and Cairo is 0,1,0,0, each city a corner of its own.
A ticket represented by concatenated one-hot vectors for channel, product and severity, forty switches, five of them on.
A vocabulary of ten thousand words, each a one-hot vector ten thousand long, which is why embeddings were invented.

Think of it this way

Think of a row of light switches, one per option, where exactly one may be on: the pattern tells you the choice instantly, and no two choices share a switch.

Worked example

from switches to positions

Suppose a support team encodes its fifty products as one-hot vectors for a routing model. The figures are illustrative.

The model works, but products launched last month have too little data to earn their switches, and their tickets misroute.

An embedding trained on ticket text places the new products near their natural neighbors, and routing works from day one on borrowed similarity.

The one-hot vector had told the truth and nothing else: each product an island, with no geography between them, until the embedding drew the map.

Illustrative case study

A fictional analytics team shuffled the category order in their one-hot vectors between training and serving, not knowing position is identity.

Every category became another category: Paris read as Cairo, and the model scored confidently wrong until a checksum on the encoder's legend caught the swap.

The invented lesson: a one-hot vector is positions, and positions are a contract. Version the legend with the model, or the switches lie.

Common misconceptions and mistakes

The 1 marks importance. It marks membership: the switch that is on, nothing more, with no magnitude or ranking in it.
The vectors carry similarity. All categories are equidistant; any real resemblance must be learned, which is what embeddings are for.
Order does not matter. Position is identity, and a shuffled vector silently renames every category.
It compresses information. It expands it: one fact spread across a column per category, which is why wide vocabularies need denser representations.

Business use

Version the category-to-position mapping with every model, and checksum it between training and serving.

Remember the geometry: one-hot vectors encode no category similarity, so do not expect neighbourly behavior.

Plan the move to embeddings when the switch space grows wide or new categories arrive often.

Read model inputs as rows of switches when debugging; most encoding bugs are visible to the naked eye.

Keep the encoder legend in the same repository as the model, versioned together, so the switches never drift from their meanings.

Check your understanding

Why do one-hot vectors encode no similarity between categories?

Answer: Because every category is placed at its own corner, equidistant from all others by construction, so the representation is honest about membership but silent about resemblance, which is why learned embeddings replace the switches when similarity matters.

Related terms

Learn next

Be Productive Using AI builds the data literacy that makes feature choices legible: https://aiforbusinessandlife.com/academy/be-productive-using-ai. [C1]

Sources

[S24] scikit-learn documentation, Glossary of Common Terms and API Elements; [S12] Aalto University, Dictionary of Machine Learning. Full URLs are in the source ledger. Verified 11 October 2026.

Previous: One-hot encoding Next: One right answer