用于独立特征学习的稀疏非负矩阵分解算法  

Sparse non-negative matrix factorization algorithm for independent feature learning

作　　者：黄卫春[1] 赵杨 熊李艳[2] Huang Weichun;Zhao Yang;Xiong Liyan(School of Software,East China Jiaotong University,Nanchang 330013,China;School of Information Engineering,East China Jiaotong University,Nanchang 330013,China)

机构地区：[1]华东交通大学软件学院,南昌330013 [2]华东交通大学信息工程学院,南昌330013

出　　处：《计算机应用研究》2020年第4期986-989,共4页Application Research of Computers

基　　金：国家自然科学基金资助项目(61562027);江西省科技厅资助项目(201616BBI90032,20161BAB212050)。

摘　　要：为了防止特征的共适应性,提出利用余弦相似度来减少潜在特征之间的相关性,从而提高非负矩阵分解(non-negative matrix decomposition,NMF)的独立特征学习能力。此外,为了使得分解后的矩阵具有较好的稀疏性,提出在传统NMF模型中引入L2,1/2稀疏约束,增强了算法的局部学习能力和稳健性。因此,潜在特征中的语义信息更加明显,潜在空间的表示更具有判别性。在fetch_20newsgroups数据集上对文档聚类的实验结果表明,提出的INMF算法在一系列评价指标上效果都优于传统的NMF、SNMF等算法模型。To prevent the co-adaptability of features,this paper proposed cosine similarity to reduce the correlation between potential features so as to improve the independent feature learning ability of NMF.In addition,in order to make the decomposed matrix have better sparsity,it also introduced L2,1/2 sparse constraint into the traditional NMF model to enhance the local learning ability and robustness of the algorithm.Therefore,the semantic information in the latent features was more obvious,and the representation of the potential space was more discriminative.The experimental results of document clustering on fetch_20newsgroups dataset show that the proposed INMF algorithm is superior to the traditional NMF and SNMF algorithm models in a series of evaluation indexes.

关 键 词：非负矩阵分解 L 2 1/2稀疏 独立特征学习 余弦相似性 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术]