rCUR: an R package for CUR matrix decomposition

András Bodor, István Csabai, Michael W. Mahoney, Norbert SolymosiBMC Bioinformatics 2012 · DOI · OpenReview

Abstract

The rCUR package introduces CUR matrix decomposition to the R ecosystem, providing a highly interpretable alternative to traditional dimensionality reduction methods like Principal Component Analysis (PCA) or Singular Value Decomposition (SVD). By representing large datasets through a small subset of actual rows and columns—rather than abstract linear combinations—rCUR enables researchers to identify influential physical variables, such as specific genes, while maintaining high classification accuracy and capturing essential biological trends.

Introduction

In fields such as genomics and astronomy, researchers frequently manage massive matrices where the number of variables and samples can reach hundreds of thousands. To interpret this "big data," dimensionality reduction techniques like PCA and SVD are standard. However, these methods often suffer from an "interpretability gap": the resulting principal components are mathematical abstractions—linear combinations of all original data points. For instance, a principal component in a gene expression study might involve 20,000 genes, making it nearly impossible for a biologist to identify the specific pathways driving the signal. The rCUR package addresses this by selecting "real" representative columns and rows that capture the essence of the entire dataset.

Methods

The rCUR package implements CUR matrix decomposition, which approximates a large matrix $A$ as the product of three smaller matrices: $A \approx CUR$. Here, C (Columns) and R (Rows) are small subsets of actual data from the original matrix, and U is a linking matrix that joins them.

The core mechanism for selecting these rows and columns is the statistical leverage score. Traditionally used in regression to identify outliers, leverage scores quantify the "influence" a specific row or column has over the best low-rank approximation of the data.

Leverage scores across gene IDs used for feature selection.

The package offers several selection strategies:

  • Randomized: The classic theoretical approach using importance sampling.
  • Top Scores: A deterministic approach picking the highest-leverage features, which improves reproducibility in biological research.
  • Ortho-top-scores: An experimental method that considers both leverage and the orthogonality of selected vectors to minimize redundancy.

Results

The authors demonstrated the package's utility using two primary gene expression datasets: a soft tissue tumor dataset (31 samples, 5,520 genes) and a global human gene expression map (5,372 samples, 22,283 genes).

Massive Compression and Trend Preservation

In the soft tissue tumor case study, rCUR selected just 27 genes—only 0.5% of the total. Despite this drastic reduction, PCA plots using only these 27 genes showed cluster separation nearly identical to the full 5,520-gene dataset, effectively distinguishing between GIST, LEIO, and SARC phenotypes.

Comparison of PCA plots using all genes (top) versus a selected subset of 27 genes (bottom).

In the larger 5,372-sample study, the authors reduced the probe count from 22,300 to 250. This 99% reduction still captured major biological trends, such as hematopoietic and malignancy axes. The study found that optimal separation performance was achieved with approximately 150 genes, suggesting that "more data" does not always translate to better classification.

Discussion

The primary advantage of rCUR is that it does not sacrifice performance for clarity. By using actual columns (e.g., specific genes), researchers can immediately identify biomarkers driving disease classification. Leverage scores, traditionally diagnostic tools, are reframed as powerful engines for feature selection.

While the package significantly accelerates exploratory data analysis, some methods like ortho.top.scores remain experimental. Additionally, the initial calculation of SVD required for leverage scores can be a bottleneck for extremely large matrices. Future iterations may benefit from integrating faster, randomized SVD algorithms to further scale the package's capabilities.

Summary updated

Abstract

Background Many methods for dimensionality reduction of large data sets such as those generated in microarray studies boil down to the Singular Value Decomposition (SVD). Although singular vectors associated with the largest singular values have strong optimality properties and can often be quite useful as a tool to summarize the data, they are linear combinations of up to all of the data points, and thus it is typically quite hard to interpret those vectors in terms of the application domain from which the data are drawn. Recently, an alternative dimensionality reduction paradigm, CUR matrix decompositions, has been proposed to address this problem and has been applied to genetic and internet data. CUR decompositions are low-rank matrix decompositions that are explicitly expressed in terms of a small number of actual columns and/or actual rows of the data matrix. Since they are constructed from actual data elements, CUR decompositions are interpretable by practitioners of the field from which the data are drawn. Results We present an implementation to perform CUR matrix decompositions, in the form of a freely available, open source R-package called rCUR. This package will help users to perform CUR-based analysis on large-scale data, such as those obtained from different high-throughput technologies, in an interactive and exploratory manner. We show two examples that illustrate how CUR-based techniques make it possible to reduce significantly the number of probes, while at the same time maintaining major trends in data and keeping the same classification accuracy. Conclusions The package rCUR provides functions for the users to perform CUR-based matrix decompositions in the R environment. In gene expression studies, it gives an additional way of analysis of differential expression and discriminant gene selection based on the use of statistical leverage scores. These scores, which have been used historically in diagnostic regression analysis to identify outliers, can be used by rCUR to identify the most informative data points with respect to which to express the remaining data points.
BibTeX
@article{bodor2012rcur,
  author = {András Bodor and István Csabai and Michael W. Mahoney and Norbert Solymosi},
  title = {{rCUR: an R package for CUR matrix decomposition}},
  journal = {BMC Bioinformatics},
  year = {2012},
  doi = {10.1186/1471-2105-13-103},
}
Figure