Abstract: Firstly, by introducing adaptive strategy, we proposed a selfadaptive nonnegative matrix factorization based on gradient descent. Secondly, by comparing the distance between reconstructed nonnegative matrix and selfadaptive regulation, the problems of 
randomness and the number of basic vectors validation for traditional nonnegative matrix factorization were solved, and the basic vectors generated by the algorithm were more representative. Finally, taking the analysis and validation of undergraduate achievement of a college of Jilin University as an example, we investigated effectiveness of the proposed algorithm. The experimental results show that  compared with the traditional nonnegative matrix method, the selfadaptive nonnegetive matrix factorization method has better robutness and reduces the error rate by 20.16%.

Key words: nonnegative matrix factorization, selfadaptive, randomness, robustness