Time-constrained NMF algorithm and its application in brain dynamic functional network dimension reduction

In order to ensure the time attributes of high-dimensional data can be preserved during the process of dimensionality reduction, a time-constrained non-negative matrix factorization (TNMF) algorithm is proposed. A time attribute constraint model is constructed by considering time series information, data dimension value, decomposition error together, then the optimal base matrix dimension value can be calculated. Using this algorithm, the spatial structure and time sequence information of original high-dimensional data will not change while the dimension reducing. The experiment results in human brain dynamic functional network dimension reduction show that this algorithm is superior to commonly used algorithms in temporal feature extraction, clustering visualization and clustering index.