- Tytuł:
- Multiplicative Algorithm for Correntropy-Based Nonnegative Matrix Factorization
- Autorzy:
-
Hosseini-Asl, E.
Zurada, J. M. - Powiązania:
- https://bibliotekanauki.pl/articles/108758.pdf
- Data publikacji:
- 2013
- Wydawca:
- Społeczna Akademia Nauk w Łodzi
- Tematy:
-
Nonnegative Matrix Factorization (NMF)
Correntropy
Multiplicative Algorithm
Document Clustering - Opis:
- Nonnegative matrix factorization (NMF) is a popular dimension reduction technique used for clustering by extracting latent features from highdimensional data and is widely used for text mining. Several optimization algorithms have been developed for NMF with different cost functions. In this paper we evaluate the correntropy similarity cost function. Correntropy is a nonlinear localized similarity measure which measures the similarity between two random variables using entropy-based criterion, and is especially robust to outliers. Some algorithms based on gradient descent have been used for correntropy cost function, but their convergence is highly dependent on proper initialization and step size and other parameter selection. The proposed general multiplicative factorization algorithm uses the gradient descent algorithm with adaptive step size to maximize the correntropy similarity between the data matrix and its factorization. After devising the algorithm, its performance has been evaluated for document clustering. Results were compared with constrained gradient descent method using steepest descent and L-BFGS methods. The simulations show that the performance of steepest descent and LBFGS convergence are highly dependent on gradient descent step size which depends on σ parameter of correntropy cost function. However, the multiplicative algorithm is shown to be less sensitive to σ parameterand yields better clustering results than other algorithms. The results demonstrate that clustering performance measured by entropy and purity improve the clustering. The multiplicative correntropy-based algorithm also shows less variation in accuracy of document clusters for variable number of clusters. The convergence of each algorithm is also investigated, and the experiments have shown that the multiplicative algorithm converges faster than L-BFGS and steepest descent method.
- Źródło:
-
Journal of Applied Computer Science Methods; 2013, 5 No. 2; 89-104
1689-9636 - Pojawia się w:
- Journal of Applied Computer Science Methods
- Dostawca treści:
- Biblioteka Nauki
Artykuł