ACF-MVSelect: An adaptive anchor learning and heterogeneous correlation fusion framework for multi-view instance selection

This paper addresses the issue of feature heterogeneity in multi-view data for machine learning instance selection. Since existing instance selection methods fail to integrate sample correlations across views, a novel feature-heterogeneous instance selection method is proposed. Based on this, an enhanced anchor learning method is proposed for multi-view data to derive view-specific representations of sample distributions. Different from alternative approaches, an adaptive mechanism based on rate-reduction theory is proposed to determine the number of anchors for distinct distributions across views, thereby preventing the failure to characterize sample distributions in some views due to an insufficient number of anchors. Leveraging these anchors, the samples are encoded to construct a multi-view sample correlation graph that captures both inter-view consistency and view-dependent similarities. Finally, an Infomap-based method is applied to identify clusters of similar samples, on which an instance selection method can select representative instances to mitigate data redundancy. Experimental results show that the proposed framework achieved 84.34% average accuracy, outperforming traditional methods by 3.74%-16.42% and multi-view methods by 2.25%-12.78%, validating its effectiveness.