应用稀疏约束的非负正则矩阵学习算法

针对非负矩阵分解后的数据稀疏性较低，训练样本偏多导致运算规模持续增大的普遍现象，本文提出基于稀疏约束的非负正则矩阵学习算法，本文算法是在样本几何结构信息条件上执行非负矩阵分解操作，并且与学习算法结合，不仅能够有效保持样本局部结构，还能够充分利用前期分解结果参加迭代运算，从而达到降低运算时间目的. 本文实验表明与其他算法比较来说，本文方法在ORL人脸数据库上最多节省时间14.84 s，在COIL20数据集上为136.1 s；而在分解后数据的稀疏性上，本文方法在ORL人脸数据库上的稀疏度提高0.0691，在COIL20数据集上为0.0587. 实验结果表明了算法有效性.

Learning Algorithm of Non-Negative Factorization Matrix with Sparseness Constraints

With the aim to enhance the sparseness of the data obtained after factorization and to improve the operational scale with the increased number of training samples of data, a learning algorithm of graph regularized non-negative matrix factorization with sparseness constraints is proposed in this study. The algorithm is based on non-negative matrix factorization and the geometry information sample above, and it also combines with the learning algorithm. It cannot only effectively maintain the local structure of the sample, but can also combine with the learning step using the result of previous factorization involved in iterative computation to reduce running time. Experimental results on both ORL and COIL20 datasets show that the non-negative factorization matrix with sparseness constraints algorithm can save 14.84 s computing time on ORL face database and 136.1 s on COIL20 dataset. Also, after the decomposition of sparse data, the non-negative factorization matrix with sparseness constraints algorithm on ORL face database can increase the sparsity up to 0.0691 and sparsity 0.0587 on COIL20 dataset. The experimental results show the effectiveness of the algorithm.