jaak-s / sparsematrixproduct

Example Java code for implementing very fast crossproduct for sparse matrixes

sparseMatrixProduct

Example Java code for implementing very fast crossproduct for sparse (boolean) matrixes.

Benchmark of crossproduct

Summary: SparseBoolMatrix is 3.5x faster compared to Matrix package in R.

Two sparse boolean matrices:

• m1 of size 204717 x 2759 with 345,434 non-zeros
• m2 of size 204717 x 4000 with 414,863 non-zeros

Calculation of crossproduct between m1 and m2 ( t(m1) %*% m2 ) using Intel i7-4600M and Ubuntu 13.10.

• SparseBoolMatrix takes 190ms (Java 7).
• R Matrix package takes 700ms (R 3.1.0, Matrix_1.1-3).

Measured Java code:

m1.prod( m2 );

Measured R code:

crossprod(m1, m2)

Notes:

• In R the matrices m1 and m2 are sparse column-oriented matrices (created using sparseMatrix).
• The Matrix package in R uses SuiteSparse under its hood, so the crossproduct is running in optimised C code.
• The measured computation time is only for crossproduct (not for loading the matrices or other stuff).