An Effective NMF-Based Method for Supervised Dimension Reduction

Ngo Van Linh, Nguyen Kim Anh, Khoat Than

Abstract: Sparse topic modeling is a potential approach to learning meaningful hidden topics from large datasets with high dimension and complex distribution. We propose a sparse NMF-based method for supervised dimension reduction which aims to detect the particular topics of each class. Beside exploiting constraint convex combination of the hidden topics for each instance, our method separably learns among classes to extract interpretable and meaningful class topics. Our experimental results showed the effectiveness of our approach via significant criteria such as separability, interpretability, sparsity and performance in classification task of large datasets with high dimension and complex distribution. Our obtained results are highly competitive with state-of-the-art NMF-based methods.