Research On Classification Algorithms Via Non-negative Matrix Factorization

Classification algorithm is the main research direction in the field of artificial intelligence and machine learning, and it is also an important method in data mining. Classification is to construct an effective classifier by the features of the sample points and category labels, and achieve the classification of unlabeled data. Non-negative matrix factorization is a new local feature selection method, which is successfully applied to the field of text mining, image recognition and speech recognition. It makes the original data matrix to the product of two matrices by alternating minimization method, and makes all matrices meet the non-negative requirements. Non-negative matrix factorization effectively mapping high-dimensional data into low dimensional space, realize the data compression and dimension reduction, improve the efficiency of the algorithm, and enhances the interpretability of data. Developing fast and efficient classification on large scale data sets via non-negative matrix factorization is high-significance.This thesis proposes a classification framework based on NMF, and extends it to semi-supervised classification and feature selection algorithm. The main idea of the algorithm is to build probability matrix with features and labels by using NMF from origin data matrix and labels matrix, then decide classifier according to the probability matrix. This paper proposes four classification algorithms via non-negative matrix factorization. They are naive NMF classification algorithm, improved NMF classification algorithm, naive NMF semi-supervised classification algorithm and improved NMF semi-supervised classification algorithm. This paper also introduces the application of a variety of NMF in classification.Finally, this paper designs a variety of experiments, and compares the classification algorithms based on NMF and classical classification algorithms. The result of experiments show that our classification framework based on NMF is superior to or close to classical algorithm, in terms of the efficiency of the algorithm is particularly outstanding. And, it can be applied to different types of datasets, especially, it can be used to solve the classification of high-dimensional and sparse data.