Start from the cost of the query
Skyline community search becomes expensive when node attributes are multi-dimensional. The research question was not simply whether a graph model could predict a useful score, but whether learned representations could reduce the work performed during online search.
Match the objective to the structure
SLGCN uses two unsupervised signals: dominance loss captures how nodes compare across attributes, while link loss preserves local structural cohesion. Neither signal is sufficient alone for the target problem.
Keep learning and search connected
The learned representation becomes useful through Skyline Community Score and Expected Score Gain. The online algorithm uses those quantities to decide how to expand a candidate community.
That connection between representation, objective, and search procedure was the most important design lesson: a model is valuable when it changes the cost or quality of the system around it.