Are items always columns?

No. This entry follows the source glossary's column convention, but row-oriented storage is common elsewhere.

Does every recommender use an item matrix?

No. It is especially associated with factorization and embedding-based designs, not a universal requirement.

Can the same scores come from different factors?

Yes. Compatible transformations of both factors can preserve their product, so latent coordinates need not be unique.

AI Glossary / Mathematics and optimization / Intermediate

Item matrix

Definition

An item matrix is a collection of learned item representations used in a matrix-factorization recommender. Each item has an embedding whose components describe latent factors learned from interaction data. In the source glossary's convention, the matrix has one column per item and one row per embedding dimension. Combined with a compatible user matrix, it produces user-item prediction scores.

What it means

A recommender may start with a large table of interactions: users along one dimension and products, articles, or videos along the other. Most possible interactions are unobserved. Matrix factorization represents users and items with shorter vectors whose combinations aim to explain the available evidence.
The item matrix is the item side of this representation. Its numbers are learned parameters, not necessarily the item's price, age, genre, or other explicit metadata. A coordinate may reflect several patterns at once and may lack a clear human label. The matrix is also distinct from the original user-item interaction table.

How it works

Let U have m user rows and k latent dimensions, and let Q have k rows and n item columns. Then UQ is an m-by-n score matrix. The score for user i and item j is the dot product of user row U_i with item column Q_j.
Training adjusts the factors according to a chosen loss, weighting scheme, and regularization. Unobserved interactions require deliberate treatment rather than an automatic interpretation as dislike. Some sources instead store item vectors as rows of an n-by-k matrix V and write UV-transpose. Both conventions are valid when dimensions, indexing, and multiplication are consistent.

Three practical examples

An article recommender can store one learned vector per article and score candidate articles for a user's representation.
A media service can retrieve candidate items using its learned item factors before applying additional ranking rules.
A recommendation analyst can inspect which items have similar representations, while testing whether that similarity is useful for the intended task.

Think of it this way

Think of each item as having a short numerical profile in a shared coordinate system, and each user as having a compatible preference profile. Combining them produces a score. The analogy is limited because latent coordinates are learned jointly and need not correspond to named preferences or objective item properties.

Worked example

Consider an original two-user, three-item example with two latent dimensions. Let U have rows (1,2) and (2,−1). Let Q have rows (1,0,2) and (0,1,1), so item A's column is (1,0), B's is (0,1), and C's is (2,1).
The first user's scores are 1 for A, 2 for B, and 4 for C. The second user's scores are 2 for A, −1 for B, and 3 for C. Thus the predicted score matrix has rows (1,2,4) and (2,−1,3). These numbers are illustrative scores, not calibrated probabilities or guaranteed star ratings.
For the first user, the dot-product ranking is C, then B, then A. If an application instead normalized both user and item vectors and used cosine similarity, its scores would be 1/sqrt(5) for A, 2/sqrt(5) for B, and 4/5 for C. B would then rank ahead of C. Normalization changes the scoring rule and can change recommendations.
The latent coordinates are not uniquely meaningful. Double the first column of U and halve the first row of Q. U becomes rows (2,2) and (4,−1), while Q becomes rows (0.5,0,1) and (0,1,1). Their product remains exactly the same. The data fit alone therefore does not justify treating the first coordinate's numerical magnitude as a fixed real-world trait.
This tiny example stores 10 factor entries versus 6 entries in the full score matrix, so it is not a storage saving. With 1,000 users, 500 items, and 20 factors, the factors contain 30,000 entries versus 500,000 dense scores. That comparison concerns dense score storage, not all costs of a real system.

Illustrative case study

In a fictional example, Willow Reads builds an article recommendation prototype. A developer labels the first latent coordinate technical depth after noticing several technical articles nearby. A later training run changes the coordinate values while maintaining similar prediction quality.
The team removes that unsupported label and documents the vectors as model-dependent representations. It evaluates recommendations directly and keeps article metadata separate for explanations. This invented case illustrates why a latent factor should not automatically become a user-facing claim about an item's meaning.

Common misconceptions and mistakes

An item matrix is not an item-by-item similarity matrix. The former stores factor coordinates; the latter stores pairwise comparisons and may have very different dimensions and storage requirements.
Do not assume every factorization has interpretable axes or produces calibrated scores. Bias terms, confidence weighting, and the objective definition also affect predictions. A negative raw score can be valid in a model even if the displayed recommendation scale is nonnegative.
New items with no interactions present a cold-start problem for a purely interaction-based factorization. Adding a new identifier does not by itself provide a well-estimated embedding. Metadata-based methods or other initialization strategies require an explicit design.

Business use

The item matrix is a practical model asset: its item identifiers, ordering, dimensions, and version must remain aligned with the scoring system. Loading factors from one training run with incompatible user factors from another can yield plausible-looking but meaningless scores.
For a business application, evaluate recommendation quality and operational constraints rather than celebrating factor compression alone. Consider catalog updates, new items, filtering rules, and the cost of scoring a large candidate set. Explain recommendations using supported evidence; a coordinate should not be presented as a customer's preference or an item's factual characteristic merely because it has a convenient name.

Check your understanding

U is 200-by-12 and Q is 12-by-80. How many items and scores per user are represented?

Answer: There are 80 item columns, and UQ produces 80 scores for each of the 200 users.

Related terms

Learn next

Recalculate the worked scores using row-oriented item storage V=Q-transpose. Then compare dot-product and cosine rankings and explain which model assumptions would justify each scoring rule. Keep item identifiers attached to every vector during the exercise.

Dot product

Sources

Primary official sources: Google Machine Learning Glossary, item matrix: https://developers.google.com/machine-learning/glossary#item-matrix ; Google recommendation course, matrix factorization: https://developers.google.com/machine-learning/recommendation/collaborative/matrix . The orientation distinction was checked against both sources. All factor values and storage calculations above are original.

Previous: Item-item collaborative filtering Next: Item response theory