SVD based initialization: A head start for nonnegative matrix factorization

By Christos Boutsidis, Efstratios Gallopoulos. Pattern Recognition, 2008

739 citations recorded by the source corpus.

Subfields: Computational Mathematics, Computational Theory and Mathematics, Signal Processing

View Paper: https://www.semanticscholar.org/paper/0bc4ffe87343f3fb37c0c5c7c2115e74490c2d05

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Computational Mathematics

In Computational Theory and Mathematics

In Signal Processing

Citations in the corpus

Song recommendation with non-negative matrix factorization and graph total variation (2016, methodology): “As our model is not jointly convex, initialization may change the performance of the system, so we use the nowadays standard technique of NNDSVD [24] to get a good approximate solution.”
The Why and How of Nonnegative Matrix Factorization (2014, background): “Boutsidis et al. [11] proposed to replace each rank-one factor in =1 + , selecting the one with larger norm and scaling it properly.”
A flexible R package for nonnegative matrix factorization (2010, methodology): “The NMF package currently implements the classical random initialization, the non-negative double singular value decomposition (NNSVD) approach from [27], as well as an initialization based on Independent Component Analysis (ICA), where the entries for W and H are set to the positive part of the…”
Data Fusion by Matrix Factorization (2013, methodology): “Other more informed initialization algorithms include random C [17], random Acol [17], non-negative double SVD and its variants [18], and k-means clustering or relaxed SVD-centroid initialization [17].”
Comment-based multi-view clustering of web 2.0 items (2014, background): “However, research on NMF have found that proper initialization plays an important role in the performance of NMF in many applications [6, 26].”
On the Complexity of Nonnegative Matrix Factorization (2007, methodology): “This observation was also used by Boutsidis and Gallopoulos [5] to develop a heuristic algorithm for NMF.”
Anderson Acceleration for Fixed-Point Iterations (2011, methodology): “2] for implementing ANNLS and SVD-based initialization of the iterations (see also Boutsidis and Gallopoulos [2]).”
Nonnegative Matrix Factorization: A Comprehensive Review (2013, methodology): “Besides, SVD or PCA has always a more compact spectrum than NMF [31].”
A Deep Matrix Factorization Method for Learning Attribute Representations (2015, methodology): “Similarly, to speed up the convergence rate of NMF we use the Non-negative Double Singular Value decomposition (NNDSVD) suggested by Boutsidis et al. [42].”
Convex and Semi-Nonnegative Matrix Factorizations (2010, background): “Next we can show that the limiting solution of the update rule of Eq. ( 26 ) satisfies”
An analysis of the coherence of descriptors in topic modeling (2015, methodology): “To address the instability introduced by random initialization in standard NMF, the deterministic NNDSVD initialization method was also employed (Boutsidis & Gallopoulos, 2008).”