AI Glossary / AI and machine learning / Advanced
Cross decomposition
Definition
Cross decomposition is a family of methods that finds hidden dimensions linking two sets of measurements, such as customer survey answers and purchase behavior, so each set can be predicted from the other.
What it means
Businesses often measure the same thing twice in different vocabularies: what people say in surveys and what they do in tills. Cross decomposition looks for the shared drivers underneath both.
The methods, partial least squares and canonical correlation analysis being the best known, extract a few latent dimensions that explain as much of the covariation between the two sets as possible.
Each dimension is a weighted mix from both sides: for example a quality-seeking dimension that loads on certain survey answers and on certain buying patterns at once.
The result is both a map and a bridge: it shows which sides of the two measurement worlds move together, and it lets you predict one set from the other.
For a business, cross decomposition is how survey scores get connected to revenue, or sensor readings to product quality, when both sides are wide tables with many correlated columns.
Partial least squares is the practical favorite because it copes when columns outnumber rows, a situation where ordinary regression fails, which is common in sensor and survey data.
How it works
Start with two tables measured on the same items: N customers, with survey answers in one table and behavior in the other.
The method finds the first pair of mixes, one per table, that correlate most strongly with each other.
It removes that shared signal and repeats, extracting further independent dimensions until the useful covariation is exhausted.
The dimensions are inspected and named from their strongest loadings on both sides, which is where interpretation happens.
The fitted model then predicts the second table for new items from the first, or reveals which columns carry the link.
Scaling is essential. Each column is standardized first, otherwise the measures in the largest units dominate the dimensions that are found.
Three practical examples
- Marketing: linking brand-survey answers to actual purchase behavior across customers.
- Manufacturing: connecting dozens of process sensor readings to a handful of final-quality measurements.
- Food science: linking tasting-panel scores to laboratory instrument readings.
Think of it this way
Think of two translators describing the same conversation, one in formal words and one in slang. Cross decomposition finds the handful of themes both are really talking about, underneath their different vocabularies.
The limit: if the two tables are not genuinely related, the method will still find the least-unrelated mixes, and a creative analyst can mistake them for meaning.
Worked example
survey answers meet till data
Suppose a retailer has 40 survey questions and 60 behavior measures for the same 5,000 customers, and suspects quality perception drives buying.
Cross decomposition extracts a first shared dimension that loads strongly on three quality questions and on premium-product purchases. The example is illustrative.
A second dimension links price questions to promotion response. The marketing team finally sees which survey movements predict till movements.
They build a small predictor: quarterly survey waves now forecast category shifts two months ahead.
Before trusting it, the team checks the dimensions on a fresh quarter of survey and till data. The first dimension still predicts premium purchases, while the second is weaker, so only the first goes into planning.
Illustrative case study
A fictional food producer struggles to link lab measurements to panel taste scores across hundreds of product variants.
Cross decomposition finds two latent dimensions that explain most of the link, and the lab learns which of its thirty instruments actually track taste.
The invented lesson: the value was not prediction but focus. Knowing which measurements carry the link cut the lab's testing panel by half.
The producer repeats the exercise when a new product line launches, since the dimensions found for the old range may not describe the new one. A quick refit and a check on held-out samples tells them whether to trust the old map.
Common misconceptions and mistakes
- Correlation between the tables proves a real link. The method finds the strongest existing covariation, which may be confounded or coincidental.
- More dimensions are better. Each extra dimension explains less and risks fitting noise; a few strong ones is the goal.
- It proves causation. It finds co-movement; claims about causes need design and domain knowledge.
Business use
Use cross decomposition when you have two wide, related tables on the same items and want the few shared drivers, not column-by-column fishing.
Insist on held-out validation before acting on the discovered dimensions, because the method will always find something.
Use the loadings, not just the predictions. The list of which survey questions and which behaviors drive each dimension is often the most actionable output.
People also ask
What is partial least squares?
A leading cross-decomposition method that extracts dimensions maximizing covariation between two tables.
How many dimensions should I keep?
As many as hold up on held-out data, usually a small number.
Is this the same as ordinary correlation?
It is correlation between learned mixes of columns, not between individual columns.
Do I need statistics training to use it?
The tools are available in common packages, but interpreting dimensions and validating them benefits from an experienced analyst.
Check your understanding
What do the dimensions found by cross decomposition represent?
Answer: Latent drivers shared by both measurement sets, each a weighted mix from both tables that captures part of their co-movement.
Related terms
- Latent variable
- Dimensionality reduction
- partial least squares
- canonical correlation
- multivariate analysis
Learn next
AI for Business Owners helps you connect the numbers you survey with the numbers you bank: https://aiforbusinessandlife.com/academy/ai-for-business-owners. [C2]
More in AI and machine learning
- Coupled pattern learner - The Coupled Pattern Learner, or CPL, is a semi-supervised information-extraction method that learns text patterns for categories and relations together.
- Crash blossom - A crash blossom is a headline or sentence whose wording invites a wildly wrong reading, named after a real headline about a musician whose career ended in a...
- Critic - In reinforcement learning, the critic is the part of the system that judges how good the current situation or action is, while another part, the actor...
- Cross-entropy method - The cross-entropy method is a sampling-based approach to optimization and rare-event estimation.
- Cross-fitting - Cross-fitting is a validation technique where predictions for every data point are made by a model that never saw that point in training, giving honest...
- Data analysis - Data analysis is the process of inspecting, cleaning and summarizing data to answer a question or support a decision, ranging from simple totals and charts...
Sources
[S26] scikit-learn documentation, Cross decomposition. Full URLs are in the source ledger. Verified 11 October 2026.
Previous: Cross-correlation matrix Next: Cross-dissolve