Does wide mean more examples than usual?

No. The term concerns model features and architecture, not the number of training rows.

Can a wide model represent interactions?

Yes, when interaction features are explicitly included. It does not discover arbitrary hidden representations merely by having many columns.

Is Wide & Deep an ensemble of separately trained models?

The original architecture jointly trains its wide and deep components. Simply averaging separate predictions is a different training arrangement.

AI Glossary / Statistics and probability / Intermediate

Wide model

Definition

A wide model is a linear model with many input features, often represented sparsely, that connect directly to an output score. It can include explicitly constructed feature combinations or transformations. In machine learning, this term commonly contrasts with a deep model, which learns through multiple hidden layers.

What it means

“Wide” describes the breadth of the feature representation rather than the number of hidden layers. An input may have thousands of possible categories while only a few are active for any one example. A model can therefore be very large in its possible inputs yet simple in how it combines them.

The score is linear in the supplied features and learned weights. However, those features can themselves be nonlinear transformations of raw inputs. A feature that becomes active only when two conditions hold lets a linear score represent that specific interaction without adding a hidden neural layer.

How it works

Choose input features such as category indicators, numeric measurements and crossed features. Represent each example as a vector x. The model forms a score z = b + Σwⱼxⱼ, where b is an intercept and each wⱼ is a learned weight. For binary classification, a logistic link can convert the score into a probability estimate.

Training adjusts weights against a loss function, usually with regularization to control complexity. In a Wide & Deep architecture, the wide component is jointly trained with a deep component. A standalone wide model is not automatically such a hybrid; the architecture and training arrangement must actually include both parts.

Three practical examples

1. Product recommendations: A retailer creates features combining a visitor's previously viewed product category with a candidate product category. The model can learn weights for frequently observed combinations and use them when ranking candidates.

2. Customer support: A routing model combines language, issue type and selected interactions. A team can inspect active features behind a score while still testing whether the routing decisions work well for uncommon cases.

3. Content ranking: A learning platform uses topic indicators and explicitly chosen context features. A wide model provides a baseline against which a more complex recommendation architecture can be evaluated on the same data.

Think of it this way

Think of a large mixing desk with many labeled switches feeding one output. Each active switch contributes according to its learned setting. A specially constructed switch can represent two conditions occurring together. The desk may have many controls, but it does not contain the chain of internal transformations associated with a deep network.

Worked example

Consider an original simplified logistic model for whether a user selects a recommended tutorial. Let x₁ indicate that the user previously viewed analytics material, and x₂ indicate that the candidate is a spreadsheet tutorial. Add the crossed feature x₃ = x₁x₂.

Suppose the fitted score is z = −1.5 + 0.6x₁ + 0.3x₂ + 1.2x₃. When x₁ = x₂ = 1, the crossed feature is also one, so z = −1.5 + 0.6 + 0.3 + 1.2 = 0.6. The logistic probability is 1/(1 + exp(−0.6)) ≈ 0.6457.

Without the crossed term, keeping all other weights unchanged for illustration, the score would be −0.6 and the probability approximately 0.3543. The difference shows the interaction's contribution in this fixed-weight calculation. It does not show what a separately retrained model would predict, establish causal user preferences or demonstrate that either probability is calibrated on real data.

Illustrative case study

In this fictional case, Fern Learning builds a wide baseline for ranking short courses. Engineers add a small set of documented topic-context crosses and compare it with a model using only individual indicators. They evaluate on later user sessions to reduce the risk of measuring memorization of repeated historical events.

Some frequent combinations improve ranking, while rare crosses behave inconsistently. The team strengthens regularization, removes unsupported features and reviews performance for new topics. It retains the wide model as a transparent comparison point when testing a hybrid architecture, rather than assuming a larger feature space must always produce better recommendations.

Common misconceptions and mistakes

A wide model is not simply a deep neural network with a large hidden layer. That is a different use of network width. Nor does “linear” mean that the raw-input relationship must be a straight line when transformed or crossed features are used.

Explicit feature interactions can become extremely numerous. Rare combinations may have too little evidence to estimate useful weights, while historical correlations can encode exposure patterns or other biases. Inspectable weights are helpful but do not make the model causal or automatically fair. A model that scores familiar combinations well can still struggle with new categories and changing behavior.

Business use

Wide models can provide useful baselines for sparse business data, especially where meaningful category combinations can be specified and monitored. They offer a relatively direct connection between feature construction and the resulting score.

Selection should still depend on predictive performance, calibration, latency, memory requirements and maintenance effort. Track feature dictionaries and preprocessing versions alongside model weights. Otherwise, a change in category encoding can silently change predictions even when the stored coefficient values have not changed. Evaluate new categories and rare combinations explicitly before operational rollout.

Check your understanding

In the example, when does x₁x₂ equal one?

Answer: Only when both binary indicators equal one.

Why can a linear score still capture the illustrated interaction?

Answer: The product is already included as an input feature, so its coefficient contributes when that joint condition holds.

Does the 0.6457 output prove a 64.57% real-world selection rate?

Answer: No. That interpretation needs calibration and evaluation on relevant data.

Related terms

Learn next

Next, build a small sparse baseline with clearly documented features and a time-appropriate evaluation split. Compare main-effect and crossed-feature versions. Then study embeddings and joint Wide & Deep training to understand what additional representations might contribute beyond explicit combinations.

Main effect

Sources

Google's Machine Learning Glossary, “wide model,” defines the sparse linear-model sense used here: https://developers.google.com/machine-learning/glossary#wide-model

Cheng and colleagues, “Wide & Deep Learning for Recommender Systems,” explains crossed features and jointly trained components: https://arxiv.org/pdf/1606.07792

The tutorial-selection calculation and Fern Learning case are original teaching illustrations. The supplied statistics category is retained, although the entry describes a machine-learning architecture.

Previous: Wide matrix Next: Widely applicable information criterion