Can one row have different category labels in different views?

Yes; that is a central feature.

Must the number of groups be fixed beforehand?

Not in the Bayesian nonparametric CrossCat formulation.

Do discovered dependencies prove causation?

No; they describe probabilistic associations under the model.

AI Glossary / Probabilistic machine learning / Advanced

Nested partition model

Definition

In the CrossCat-style setting used here, a nested partition model groups the columns of a data table into views and assigns a separate partition of the rows within each view. Different sets of variables can therefore support different groupings of the same entities. It allows different variable groups to support different row clusters.

What it means

A table might describe product dimensions, materials, ordering patterns, and delivery characteristics. Two products can be similar in their physical properties but very different in how customers order them. One global clustering may obscure this.
A nested partition can place related variables together, then group products differently for each variable group. Each row can have different cluster assignments across views. The word nested describes the organization of these partitions. It does not mean that all row groups form a single tree in which every smaller group sits inside one common larger group.

How it works

The outer partition assigns each column to a view. Within each view, an inner mixture model partitions rows into categories. Each category contains probability models appropriate to its columns, such as Bernoulli distributions for binary variables or suitable numerical distributions for continuous measurements.
CrossCat uses Bayesian nonparametric constructions so the numbers of views and categories need not be fixed in advance. Approximate posterior inference explores alternative structures and parameters supported by the data and priors. Predictions can average over these alternatives instead of declaring one partition unquestionably correct.
Within a specified view, variables can be dependent because they share an uncertain category assignment. Conditional component models may factor across columns given that category. Different views factor independently within a fixed model structure. Posterior uncertainty about structure must still be considered when summarizing relationships. The inferred groups describe statistical patterns, not causal mechanisms.

Three practical examples

An inventory analyst studies a mixed table of physical specifications and ordering behavior. Different variable groups may support distinct product clusters, allowing operational patterns to be examined without assuming they follow engineering categories.
A software service studies account features and usage measurements. One view could capture interaction patterns while another captures technical configuration. A customer can resemble different peers under these different descriptions, subject to validation and appropriate data handling.
An equipment study combines sensor summaries, maintenance records, and component attributes. Separate views can express different statistical signals. Persistent time trends or dependencies between successive records require additional modeling rather than being automatically explained by a table-based clustering model.

Think of it this way

Imagine maintaining several filing systems for the same collection of objects: one by physical characteristics and another by how they are used. The model learns which attributes belong together in each filing system and how to organize the objects within it. It does not insist that one set of folders answers every question.

Worked example

Consider four rows and four binary columns A, B, C, and D. Their values are: row 1 = (0, 0, 0, 0); row 2 = (0, 0, 1, 1); row 3 = (1, 1, 0, 0); row 4 = (1, 1, 1, 1).
A candidate column partition has view 1 containing A and B, and view 2 containing C and D. In view 1, rows 1 and 2 group together, as do rows 3 and 4. In view 2, rows 1 and 3 group together, as do rows 2 and 4. These partitions cross. They need not be reconciled into a single two-cluster assignment for all columns.
For prediction under this structure, stipulate two equally probable categories per view. Each binary variable has probability 0.1 of being one in the low category and 0.9 in the high category, with conditional independence inside each category. These probabilities are teaching assumptions, not estimates calculated from the four rows.
If a new row has A = 1, the high-category probability in view 1 becomes (0.5 × 0.9)/(0.5 × 0.9 + 0.5 × 0.1) = 0.9. Its predicted probability that B = 1 is then 0.9 × 0.9 + 0.1 × 0.1 = 0.82. Under this fixed independent-view model, observing A does not change C's probability, which remains 0.5 × 0.9 + 0.5 × 0.1 = 0.5.
The calculation illustrates dependence within a view and independence across specified views. It is not full CrossCat posterior prediction, which also accounts for uncertainty about partitions, parameters, and possible categories.

Illustrative case study

Fictional company Birch Inventory proposes a nested partition analysis of its product table. Its analysts compare it with a single-clustering model and reserve observed entries for predictive evaluation. They inspect whether recurring views correspond to interpretable groups of variables and whether their structure changes across posterior samples.
They also distinguish genuinely missing measurements from fields that are inapplicable to particular product types. Treating both as interchangeable missing values could distort interpretation. This invented scenario outlines a study; it does not claim that the inferred groups improved purchasing, reduced inventory, or identified causal drivers.

Common misconceptions and mistakes

The number of views is not the same as the number of row categories. Each view has its own category structure. Columns grouped together are not necessarily identical or strongly linearly correlated; their joint distribution can involve more complicated relationships.
Bayesian nonparametric does not mean assumption-free or computationally effortless. Likelihoods, priors, inference settings, and missing-data assumptions still matter. A single displayed partition is also not a complete account of posterior uncertainty. Finally, related uses of “nested partition” elsewhere in statistics may describe different constructions and should not be substituted without checking context.

Business use

Use this approach when a heterogeneous table appears to contain several statistical signals and one segmentation is too restrictive. Evaluate predictive usefulness, stability across inference runs, and the clarity of conclusions for the intended decision. Keep uncertainty visible when assigning business labels to groups, and test whether the data's sampling and missingness patterns support the interpretation.

Check your understanding

Which rows group together under columns A and B?

Answer: Rows 1-2 and rows 3-4.

Does that also describe the C - D grouping?

Answer: No.

Why does observing A leave C unchanged in the predictive illustration?

Answer: The stipulated fixed model separates them into independent views.

Related terms

Biclustering
Posterior predictive distribution
CrossCat
Bayesian nonparametrics
Dirichlet process mixture
latent category
heterogeneous data

Learn next

Compare a single mixture over all columns with separate mixtures over views, then examine how posterior samples express uncertainty about the column partition.

Sources

Murphy, PML Introduction contents, Section 21.6.2 (Crosscat): https://probml.github.io/pml-book/toc1.pdf . Mansinghka et al., CrossCat (2016), generative process consulted: https://www.jmlr.org/papers/volume17/11-392/11-392.pdf . The binary-table example and Birch Inventory scenario are original illustrations.

Previous: Nested Loop Next: Nested Relation