Description
Cluster the rows and the columns of the same data matrix separately, then reorder both by their tree orders to reveal blocks of objects sharing attributes.
A one-way cluster analysis groups either the objects or the attributes. Two-way rearrangement does both: a Q-analysis clusters the columns, an R-analysis clusters the rows, and the matrix is reordered so rows and columns follow their dendrogram orders. Blocks of 1s, or of high values, then show which objects go with which attributes, a structure that is invisible in the raw matrix even when it is small. In industrial engineering this forms machine cells for families of components; special-purpose algorithms such as bond energy and rank order clustering reach the same block form directly. The choice of similarity coefficient matters: when shared absence is meaningful, use one that counts 0-0 matches, such as simple matching.
Attributes
- Discipline
- Statistics & Modelling
- Loss type
- Unmigrated
- Stage
- Discovery, Definition
- Data
- Quantitative
- Effort
- Days
Steps
- Build the matrix, for example users by features used or tasks by screens visited, coded 0/1 or quantitatively.
- Choose a similarity coefficient; use one that counts 0-0 matches (simple matching) when shared absence is informative, Jaccard when it is not.
- Cluster the columns (Q-analysis), for example with UPGMA, and cluster the rows (R-analysis) separately.
- Reorder columns and rows into their tree orders, or apply a bond energy or rank order algorithm directly.
- Read off the blocks of co-occurring rows and columns and list the exceptions that fall outside any block.
- Interpret blocks as candidate segments, modules or feature bundles and check them against qualitative evidence.
Sources
- Cluster Analysis for Researchers (Romesburg, 1984)
Also known as
- Two-way cluster rearrangement of a binary matrix
- Q and R analysis together
- Doubly rearranged data matrix
- Block clustering
- Bond energy algorithm
- Rank order cluster algorithm
- Group technology cells
Categories
- Analytics & Behaviour
- Statistics & Modelling