Robust collaborative clustering approach with adaptive local structure learning

Highlights：

• We propose a new adaptive local structure learning strategy, which can assign the most suitable number of neighbors to each data point based on its unique spatial structure. Moreover, it is a general adaptive learning strategy that can be extensively applied to other machine learning approaches based on graph learning.

• We relax the orthogonality of the factor matrix as a regularization term, which helps reduce the burden of model solving. Furthermore, the existence of the L1 loss function allows the model to maintain good numerical performance in most noise tasks.

• We integrate the adaptive local structure learning strategy and NMTF, which enables the model to dynamically mine the local structure of the data space while performing the matrix collaborative factorization, so as to obtain a “high-quality” affinity matrix.