Ensemble Biclustering for Classification

A python implementation of the Ensemble Biclustering for Classification (EBC) algorithm. Although having "biclustering" in its name, EBC is a co-clustering algorithm that allows you to perform co-clustering on very large sparse N-dimensional matrix. For details and examples of using EBC please reference this paper.

Files

ebc.py: an implementation of EBC algorithm using sparse matrix.
ebc2d.py: a vectorized implementation of EBC algorithm using (2D) numpy dense array.
matrix.py: a N-dimensional sparse matrix implementation using python dict that allows basic get/set/sum operations.
ebc-sample-matrices/: this directory contains many example data that could be used to run EBC on.

Usage

General Usage

Using ebc is easy. First you need a sparse matrix constructed as a SparseMatrix instance defined in matrix.py. You can use matrix.py's built in read_data() method to construct the sparse matrix.

Once you have the sparse matrix and import the ebc module, you can simply do the following:

ebc = EBC(sparse_matrix, n_clusters=[30, 125], max_iterations=10)
cXY, objective, iter = ebc.run()

The returned cXY contains the clustering assignments along each axis in a list of lists, objective contains the final objective function value, iter contains the iterations that it ran for to converge.

Efficiency Considerations

In short, EBC is built for highly sparse multi-dimensional matrix, while EBC2D is a vectorized version of EBC, and is built for less sparse 2D matrix. The running time of EBC increases linearly with the number of non-zero elements, while the running time of EBC2D will only increase when matrix size increases.

If your input matrix is highly sparse or you need support for N-dimensional matrix with N > 2, you should use EBC class in ebc.py.
If your input matrix is not very sparse (e.g. 5% density), and can fit into memory with numpy, you can choose to use EBC2D class in ebc2d.py. Note that EBC2D only supports 2D matrix, since large multi-dimensional dense matrix can hardly fit into memory.

To give you more sense of the efficiency, the EBC implementation runs for ~50 seconds on our 14052 x 7272 example matrix (which is >99% sparse) on a MacBook, and in this case it is faster than EBC2D. However, in our testing, for a 5000 x 5000 matrix of 95% sparsity, EBC2D can get a x5 speedup over EBC. And this speedup grows as the sparsity decreases.

Dependencies

To run this implementation of EBC algorithm, you need to have numpy installed.

References

Percha, Bethany, and Russ B. Altman. "Learning the Structure of Biomedical Relationships from Unstructured Text." PLoS computational biology 11.7 (2015).
Dhillon, Inderjit S., Subramanyam Mallela, and Dharmendra S. Modha. "Information-theoretic co-clustering." Proceedings of the ninth ACM SIGKDD international conference on Knowledge discovery and data mining. ACM, 2003.

Questions?

This code was written by Beth Percha and Yuhao Zhang. We welcome any questions or comments, and would appreciate it if you would let us know if you make any substantial modifications or improvements. You can reach us at blpercha@stanford.edu.

blpercha/ebc