Java library for performing non-negative matrix factorization. A non-negative matrix X is factorized into the product of two matrices W and H, such that the distance between X and WH is minimal. The distance can be evaluated using the euclidean distance or the generalized Kullback-Leibler divergence with optional regularization terms. Two types of gradient-descent update rules are supported.
- Euclidean distance || X − WH ||2.
-
Multiplicative update rules for Euclidean and Kullback-Leibler distances with regularization terms. Based on D. Lee and H. Seung, Algorithms for Non-negative Matrix Factorization.
-
Fast-gradient-descent update rules for Euclidean and Kullback-Leibler distance with regularization terms. Based on N. Guan et al., Non-Negative Patch Alignment Framework.
-
Alternating Least Squares (ALS) method for solving the non-negative matrix factorization (NMF). Based on H Kim and H, Park, Nonnegative matrix factorization based on alternating nonnegativity constrained least squares and active set method.
-
Non-negative matrix factorization (NMF), performed by alternating updates of matrices W and H to minimize the distance between X and WH.
-
Active set method for solving non-negative least squares problem. Based on R. Bro and S.D. Jong, A fast non‐negativity‐constrained least squares algorithm.
-
Non-negative optimization, performed by updating matrix H to minimize the distance between X and WH.
-
Non-negative singular value decomposition (NNDSVD), used to initialize matrix W and H. Based on C. Boutsidis and E. Gallopoulos, SVD based initialization: A head start for nonnegative matrix factorization.
These instructions will get you a copy of the project up and running on your local machine for development and testing purposes.
To install the package, clone this project and use maven to build it:
git clone https://github.com/du-lab/nonnegative_matrix_factorization.git
cd nonnegative_matrix_factorization/
mvn clean install
Example: Given matrixX and num_components, perform non-negative matrix factorization using the euclidean distance with regularization, multiplicative update rule,
and NNDSVD-initialization.
final int num_points = matrixX.rows;
final int num_vectors = matrixX.columns;
// Create matrices W and H
DMatrixRMaj matrixW = new DMatrixRMaj(num_points, num_components);
DMatrixRMaj matrixH = new DMatrixRMaj(num_components, num_vectors);
// Initialize matrices W and H by the NNDSVD-method
new SingularValueDecomposition(matrixX).decompose(matrixW, matrixH);
// Choose update rules for matrices W and H:
// Multiplicative update rule for the euclidean distance with l1-regularization
UpdateRule updateRuleW = new MUpdateRule(1.0, 0.0);
// Multiplicative update rule for the euclidean distance with l2-regularization
UpdateRule updateRuleH = new MUpdateRule(0.0, 1.0);
// Perform factorization
new MatrixFactorization(updateRuleW, updateRuleH, 1e-4, 10000).execute(matrixX, matrixW, matrixH);Or, perform non-negative matrix factorization using the alternating least squares method.
final int num_points = matrixX.rows;
final int num_vectors = matrixX.columns;
// Randomly initialize matrices W and H
Random random = new Random(0);
DMatrixRMaj matrixW = rectangle(num_points, num_components, 0.0, 1.0, random);
DMatrixRMaj matrixH = rectangle(num_components, num_vectors, 0.0, 1.0, random);
// Perform factorization
new AlternatingLeastSquaresMatrixFactorization(1e-4, 10000).solve(matrixX, matrixW, matrixH);Detailed API documentation can be found here.
Code contributions are welcome. Please, contact us if you have any questions.
- Aleksandr Smirnov - Initial work - https://github.com/asmirn1
This project is licenced under the GNU GPL v2 licence - see the LICENSE file for details.
- Set a limit on the number of iterations in NonNegativeLeastSquares algorithm
- Fix the error when the passive set is empty and vector S is calculated
- Adds the active set method for solving Non-Negative Least Squares problem
- Adds the alternating least squares method for solving Non-Negative Least Squares problem
- Replaces matrix library jBlas with EJML (Efficient Java Matrix Library)