MDGFM combines feature alignment, adaptive weighting of attributes and topology, and graph refinement during self-supervised pre-training. A dual-prompt adaptation stage transfers the resulting representations to unseen domains with few labels. The work studies both transfer performance and resilience to unreliable graph connections.

LatGRL examines semantic heterophily in heterogeneous graphs. It mines latent graphs using structural and attribute similarities, then combines low- and high-frequency information to adapt to individual nodes. The framework learns without labels and includes a scalable implementation for larger graphs.

InfoMGF studies multiplex graphs whose different views contain useful information beyond their common signal. It learns refined graph structures and a fused graph, balancing shared and view-specific information while reducing irrelevant connections. The work evaluates the learned graph through downstream tasks and robustness experiments.

BMGC addresses unequal view quality in multi-relational graph clustering. An unsupervised procedure identifies a dominant view during training, and two complementary signals guide representation learning. A graph-quality metric connects structural differences to clustering behavior, with analysis on real and synthetic graphs.