[Paper Review] Turning CARTwheels: An Alternating Algorithm for Mining Redescriptions

Deept Kumar, Naren Ramakrishnan+2 more|ArXiv.org|Nov 27, 2003

Data Mining Algorithms and Applications16 references11 citations

TL;DR

This paper introduces CARTwheels, an alternating algorithm that grows two classification trees in opposite directions to mine redescriptions—equivalent or near-equivalent set-theoretic relationships between data subsets described using different vocabularies. The method enables discovery of cryptic, high-level patterns across diverse descriptors, demonstrated effectively in bioinformatics with a 10-gene circular redescription chain and improved Jaccard coefficients up to 1.0.

ABSTRACT

We present an unusual algorithm involving classification trees where two trees are grown in opposite directions so that they are matched at their leaves. This approach finds application in a new data mining task we formulate, called "redescription mining". A redescription is a shift-of-vocabulary, or a different way of communicating information about a given subset of data; the goal of redescription mining is to find subsets of data that afford multiple descriptions. We highlight the importance of this problem in domains such as bioinformatics, which exhibit an underlying richness and diversity of data descriptors (e.g., genes can be studied in a variety of ways). Our approach helps integrate multiple forms of characterizing datasets, situates the knowledge gained from one dataset in the context of others, and harnesses high-level abstractions for uncovering cryptic and subtle features of data. Algorithm design decisions, implementation details, and experimental results are presented.

Motivation & Objective

• To address the challenge of integrating multiple, diverse data descriptors in high-throughput scientific domains like bioinformatics.
• To formulate redescription mining as a novel data mining task that discovers equivalence relationships between subsets defined via different vocabularies.
• To develop an efficient, tree-based algorithm that automatically identifies set-theoretic expressions (e.g., intersections, differences) with high Jaccard similarity.
• To enable constructive induction by generating new, meaningful features that bridge conceptual and descriptive gaps in data.
• To support knowledge integration across datasets by identifying shared object subsets described in different ways.

Proposed method

• CARTwheels alternates between growing two classification trees: one from the X-descriptor side and one from the Y-descriptor side, with leaves matched to identify potential redescriptions.
• The algorithm uses a symmetric Jaccard coefficient as the primary evaluation metric: $ \text{Jaccard}(E,F) = \frac{|E \cap F|}{|E \cup F|} $, where E and F are set-theoretic expressions over descriptors.
• It applies syntactic bias to constrain expression complexity (e.g., intersections of two descriptors from X, unions or differences from Y) to ensure tractability.
• The method leverages entropy and decision tree induction principles to guide tree growth toward high-similarity leaf matches.
• It supports the use of set operations (union, intersection, difference) to refine and strengthen redescriptions beyond simple descriptor matches.
• The algorithm is designed to handle both exact and approximate redescriptions, with a focus on identifying biologically or scientifically meaningful equivalences.

Experimental results

Research questions

• RQ1 How can we automatically discover equivalent or near-equivalent subsets of data described using different vocabularies?
• RQ2 What algorithmic approach enables efficient mining of set-theoretic relationships across multiple descriptor families?
• RQ3 How can we evaluate and rank redescriptions based on their strength and biological relevance?
• RQ4 Can alternating tree induction effectively uncover subtle, high-level patterns in complex, multi-descriptor datasets?
• RQ5 To what extent can redescriptions bridge abstraction levels in data, such as from genes to pathways or biological processes?

Key findings

• CARTwheels successfully discovered a circular redescription chain (R7) involving 10 genes across three experimental comparisons, indicating coordinated biological activity under stress.
• The algorithm achieved a perfect Jaccard coefficient of 1.0 in a refined redescription: ‘Countries with declared nuclear arsenals’ ∩ ‘Defense budget > $30B’ ⇔ ‘Permanent members of U.N. Security Council’ – ‘Communist history countries’.
• A redescription involving two descriptors from X and one from Y achieved a Jaccard coefficient of 0.625, improving from an initial 0.428 through descriptor refinement.
• The method uncovered a 7-gene redescription in a real biological dataset, demonstrating its utility in identifying functionally coherent gene sets.
• The algorithm effectively identified a 3-country match (U.S.A., U.K., France) in a complex redescription involving intersection and difference operations.
• The approach demonstrated strong performance in integrating heterogeneous descriptors in bioinformatics, revealing biologically meaningful, non-obvious relationships across data abstractions.