Uncertainty-Aware Global-View Reconstruction for Multi-View Multi-Label Feature Selection

TL;DR

This work tackles feature selection in multi-view multi-label learning by introducing Uncertainty-aware Global-view Reconstruction for MVML Feature Selection (UGRFS). It unifies global-view reconstruction with an uncertainty-aware objective, learning a data-driven global-view distribution $D$ while accounting for sample confidence $C$ and view relationships to fuse information across views. The method combines a graph-based view fusion, a nonlinear mapping of the label space via $\rho(Y)$, and a group-sparsity regularizer to obtain robust, view-aware feature weights. Empirical results across six MVML datasets show that UGRFS consistently outperforms baselines and that its components (sample confidence, view-aware reconstruction) contribute meaningfully to performance and convergence.

Abstract

In recent years, multi-view multi-label learning (MVML) has gained popularity due to its close resemblance to real-world scenarios. However, the challenge of selecting informative features to ensure both performance and efficiency remains a significant question in MVML. Existing methods often extract information separately from the consistency part and the complementary part, which may result in noise due to unclear segmentation. In this paper, we propose a unified model constructed from the perspective of global-view reconstruction. Additionally, while feature selection methods can discern the importance of features, they typically overlook the uncertainty of samples, which is prevalent in realistic scenarios. To address this, we incorporate the perception of sample uncertainty during the reconstruction process to enhance trustworthiness. Thus, the global-view is reconstructed through the graph structure between samples, sample confidence, and the view relationship. The accurate mapping is established between the reconstructed view and the label matrix. Experimental results demonstrate the superior performance of our method on multi-view datasets.

Figures (5)

• Figure 1: Four typical types for the relationship between features and labels in MVML, namely, a) common-based mapping including $\textcircled{1}$ and $\textcircled{4}$, b) label consistency mapping including $\textcircled{2}$, c) view-specific mapping including $\textcircled{3}$, and d) concatenating view mapping including $\textcircled{4}$.
• Figure 2: Seven methods on SCENE in terms of Average Precision, Coverage, Hamming Loss and Ranking loss.
• Figure 3: Parameter sensitivity studies on the MIRFlickr dataset.
• Figure 4: Heatmap of the (a) sample confidence and (b) original features and uncertainty-aware features on yeast.
• Figure 5: Convergence curves analysis of UGRFS on (a) MIRFlickr and (b) VOC07.