Bridging Local Geometries via Adaptive Convex Multi-graph Fusion for Unsupervised Manifold Representation
Abstract
The quality of unsupervised manifold learning relies fundamentally on the underlying neighborhood graph. Standard linear methods utilizing a single affinity matrix often fail to capture the multiscale complexity of high-dimensional data. While multi-view algorithms integrate complementary information, additional views may be unavailable or costly to acquire in some practical scenarios. Consequently, effectively exploiting the multi-scale structural information within a single reliable view remains an important research problem. To this end, we propose Multi-Affinity Matrix Fusion-based Locality Preserving Projections (MAMF-LPP). This framework bridges fragmented local geometries through a unified joint optimization paradigm, dynamically coupling adaptive convex multi-graph fusion with manifold projection. To optimize the resulting non-convex objective, we develop an efficient Block Coordinate Descent (BCD) algorithm featuring exact row-wise probabilistic simplex projections, thereby eliminating standard heuristic approximation errors. Theoretical and empirical analysis establishes the monotonic descent of the objective, while empirical results demonstrate rapid convergence. Across multiple benchmark datasets, MAMF-LPP achieves competitive performance against representative single-graph, nonlinear, and joint-learning baselines. A case study further shows that exploiting multi-scale intra-view geometries can yield performance competitive with selected multi-view configurations under the evaluated protocol.