What counts as a view?

A subset of features that can predict the label on its own, such as text versus images, or behavior versus demographics.

How is this different from ensemble methods?

Ensembles combine models for one prediction; co-training has models teach each other new labeled data.

Is co-training still used?

The core idea lives on in modern multi-view and self-training methods, though plain co-training is rare.

When does co-training fail?

When the views are correlated, the seed labels are unrepresentative, or confident errors are allowed to compound without audits.

AI Glossary / AI and machine learning / Advanced

Co-training

Definition

Co-training is a semi-supervised method where two models learn the same task from two different views of the data, each labeling examples it is confident about to teach the other.

What it means

Some data splits naturally into two views. A web page has its text and the text of links pointing to it. A property has its listing description and its photos. Either view alone can predict the answer.

Co-training exploits this: train model A on view one and model B on view two, each starting from the few labeled examples you have.

Then let them teach. Model A labels the unlabeled examples it is most confident about and hands them to B as new training data; B does the same for A. Round by round, both improve.

The method works when the views are genuinely independent given the class, so one model's confident picks carry information the other cannot already see.

For a business, co-training is a way to stretch a small labeling budget when your data has two natural sides, though the setup is fiddly enough that it suits teams with real ML support.

Co-training is the ancestor of a wider family: self-training, where one model teaches itself, and multi-view methods in modern deep learning share its core bet that agreement across views signals truth.

How it works

Split each item's features into two views that can each, in principle, predict the label, such as description text and photo features.

Train a model on each view using the small labeled set available.

Each model labels the unlabeled items it is most confident about, and those items join the other model's training set.

Repeat for several rounds, growing both training sets, until gains stall or a round limit is reached.

Evaluate on a properly held-out labeled set, because the models' mutual confidence can inflate apparent accuracy if validation leaks into the teaching loop.

Three practical examples

Classifying web pages using page text and incoming-link text as the two views, the original application.
Product categorization using listing descriptions and image features.
Property valuation using listing text and location data, each view teaching the other.

Think of it this way

Think of two specialists examining the same patients: one reads the scans, the other the notes. Each tells the other about the cases they are sure of, and both get better without needing a senior doctor to review everyone.

The limit: if both specialists trained in the same hospital with the same habits, they mostly teach each other what they already agree on. Independence of views is everything.

Worked example

stretching 200 labels across two views

Suppose a marketplace wants to classify 50,000 listings into quality tiers but has only 200 labeled examples.

It trains one model on listing text and another on photo features. Each round, each model contributes its 50 most confident new labels to the other's training set. The figures are illustrative.

After five rounds, both models improve markedly on held-out data, though gains flatten as the easy unlabeled items run out.

The team keeps a hand-labeled audit sample outside the loop, which catches one category where both models confidently agree and are both wrong.

Illustrative case study

A fictional property portal uses co-training to value listings, text in one view and location features in the other.

It works until a market shift makes location patterns from the teaching rounds stale, and both models reinforce the outdated view. A scheduled audit against fresh sales exposes the drift.

The invented lesson: co-training amplifies what the views agree on, including outdated agreement, so independent validation must run on a schedule.

The portal now retires teaching loops after a fixed number of rounds and re-seeds with fresh labels each quarter, on the principle that mutual teaching has a shelf life like everything else it learned from.

Common misconceptions and mistakes

Co-training needs no labels at all. It needs a seed of labeled examples to start both views.
Any two feature sets work as views. The views must each predict the label and be reasonably independent given it.
Agreement between the models proves correctness. They can agree and be wrong together, which is why held-out audits exist.

Business use

Consider co-training when you have two natural views of each item and expensive labels; keep an untouched labeled audit set outside the teaching loop.

Ask vendors claiming semi-supervised magic what their two views are; without genuine view independence, co-training is relabeling its own guesses.

The hidden cost is engineering: two pipelines, two models and a teaching loop to maintain, so weigh it against simply buying more labels.

Check your understanding

What two conditions must the views satisfy for co-training to work?

Answer: Each view must be able to predict the label by itself, and the views must be reasonably independent given the label, so each model teaches the other something new.

Related terms

Learn next

AI for Business Owners explains how to judge data-saving techniques and their hidden validation costs: https://aiforbusinessandlife.com/academy/ai-for-business-owners. [C2]

Sources

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

Previous: Co-occurrence network Next: Coalescence Rule