Abstract: To address the multi-view data clustering for the multi-media data with various of feature representations as input, this paper proposes a Multi-view Regularized Matrix Factorization（MRMF） algorithm based on the well-known Nonnegative Matrix Factorization（NMF）. In detail, the proposed MRMF extends the conventional NMF to its multi-view version which considers multiple matrices as input, and then introduces the [L2,1] norm to measure the loss of matrix approximation. Furthermore, the multi-view manifold regularization is also considered to regularize the proposed matrix factori-
zation. Compared with some existing data clustering as well as multi-view clustering algorithms, the proposed MRMF is less sensitive to noises in the original data and can also better balance the importance of each view by maintaining a set of learnable weights for each view in the manifold regularization. Encouraging experimental results on numerous public multi-view datasets demonstrate the superiority of the model compared to some state-of-the-art methods.

Key words: nonnegative matrix factorization, multi-view learning, data clustering, manifold regularization

摘要： 为了解决具有多种特征属性的多媒体数据（多视图数据）挖掘问题，在非负矩阵分解（NMF）算法的基础上，提出了一种多视图正则化矩阵分解算法（MRMF），该算法使用了多元非负矩阵分解技术，同时使用[L2,1]范数描述矩阵分解的损失函数，并采用多视图流形正则化对矩阵分解进行正则化约束。与现有的一些数据聚类或多视图聚类算法相比，提出的MRMF算法不易受到原始数据中噪声的影响，而且能够充分考虑到不同视图在聚类中所具有不同权重的问题，能够对多视图数据进行较为准确的聚类。MRMF算法的有效性在一些经典的公开数据集上进行了验证，并取得了较好的聚类精度。

关键词: 非负矩阵分解, 多视图学习, 数据聚类, 流形正则化