Dictionary-Based Attention for Hyperedge Reweighting

Source abstract

We introduce Dictionary-based Attention (DA), a label-free, support-preserving block that reweights existing vertex–hyperedge memberships using dictionary-code similarity. We evaluate local dictionaries with closed-form hypergraph learning (DA-HL) on visual data and a separate shared-dictionary variant with a static incidence network on citation data. Supplementary experiments under an explicitly reconstructed visual protocol cover nine dataset–generator combinations. Four DA-versus-uniform comparisons reach the 0.05 threshold after Holm correction; the three clustering settings show no mean advantage under the original stopping rule with an iteration cap of 200. In paired clustering runs extended to 5000 iterations, mean accuracy is practically stable between 4000 and 5000 iterations within a prespecified one-percentage-point margin, but weights, propagation operators, and some predictions continue to change. This does not establish convergence or equivalence to the original 12-iteration budget. In transductive citation experiments, test features participate in shared-dictionary fitting; the paired static-backbone results are negative on Cora, close to zero on Citeseer, and slightly positive on Pubmed, with only Pubmed meeting the multiplicity-corrected threshold. Additional controls separate injected-membership ranking from downstream classification and show sensitivity to weight mapping and member-specific assignment without identifying a unique gain mechanism. The evidence supports benefits in some tested visual configurations, rather than broad effectiveness across hypergraph learning.