Semisupervised Nonnegative Matrix Factorization for Learning the Semantics

In real world there are a lot of unlabeled data, and relatively few labeled data. Unlabeled data help to learn a statistical model that can fully describe the global property of data, while labeled data help to minimize the gap between the statistical property and human beings' perception, i.e. labeled data can help to learn the semantics. Nonnegative Matrix Factorization is a popular technique in data analysis, since a lot of real world data are nonnegative. However, traditional NMF is an unsupervised learning algorithm, which means that it cannot make use of the label information. To enable NMF to make use of both labeled and unlabeled data samples, we propose a novel semisupervised Nonnegative Matrix Factorization technique for learning the semantics. The proposed algorithm extracts prior information from the labeled data, and then uses it to guide the later processing. Experimental results with different settings prove the efficacy of the proposed algorithm.