Effective feature-sample co-clustering by adaptive feature-sample co-weighting

Clustering has been a hot research topic in machine learning and data mining, and its main target is grouping data points into respective clusters according to their underlying semantic similarities. To explore more meaningful task-specific details, not only the clustering results of samples but also those of features are expected to be achieved simultaneously, leading to the co-clustering paradigm. However, existing co-clustering methods overlook the common sense that different features and samples contribute differentially to clustering results. In this work, an Effective Co-clustering model by adaptive Feature-Sample co-Weighting (ECFSW) is proposed by introducing two quantitative descriptors to characterize the different contributions of different features and samples in clustering model learning, which respectively aim at improving the model's discriminative ability and robustness. An efficient optimization algorithm is proposed to solve the ECFSW objective function in which the two descriptors are jointly optimized with the feature and sample cluster indicator matrices. Extensive experiments to evaluate the effectiveness of ECFSW are performed on both synthetic and real-world data sets from diverse aspects, including the visualization of feature and sample importance, co-clustering effect, data reconstruction ability, and clustering performance. The obtained results demonstrate the competitive performance of ECFSW compared with state-of-the-art methods.